The TriadCity Developers' Blog

I remember distinctly my first encounter with computer programming. It was 1975 and it was a simple space-conquest game the kids played at school. Perhaps an Altair, perhaps running Altair BASIC. I wanted to add some rules so I opened the source and typed them in natural language. Of course the compiler barfed, and the lesson I drew from that experience was extremely to the point: "I thought computers were supposed to be smart. This is the stupidest thing I've ever seen."

It was a naive opinion but entirely right.

Now for twenty years I've programmed highly advanced "artificial intelligence" driving the complex, rich, imaginary world TriadCity. It's still the stupidest thing I've ever seen. It's brilliant and nuanced and entirely fake.

Humanity need no fear of computer "intelligence". The proper fear is computer stupidity, paired with the hubristic self-congratulation of programmers and architects and managers and defense officials and generals who genuinely imagine themselves capable of forecasting every conceivable scenario and use case and edge case and intersection. Where the end of the species comes not at the hand of Skynet, but of Dunning-Kruger.

TriadCity relies on a lot of AI. AI-driven characters have employers to report to, homes to sweep up, meals to cook, hobbies to pursue, dates to go out on. Sports leagues play games which simulate real life; AI-driven commodity speculators calculate optimal bids. It’s a lot of AI. The old-school way to scale would be to verticalize the game server: more CPUs, more RAM, bigger-faster-more. But we’re a tiny little struggling game company and we want to keep our costs as close to nil as possible. Besides, brute force seems really kinda conceptually offensive for tasks like these.

There’s a better way. We can borrow cycles from players’ computers by offloading portions of these computations to our TriadCity clients. It’s a form of matrix computing which distributes our scaling problem to many connected machines.

We do a lot of this. I’ll note two contexts here which are interesting to me.

NLP is handled by the client. Players type natural English: “Show me what’s inside the red bag”; or, “What’s in the bag?” Client code parses those inputs before transmitting them to the server, translating them to canonical MUD-style commands based on old-timey DIKUmud. In this example, the client sends “l in red bag” or “l in bag”. Offloading this computation is very helpful to keeping the server computer small. Individually these computations aren’t that taxing, but when you scale them to tens of thousands per second you’re talking nontrivial horsepower. Pushing that CPU cost to the client is a simple form of matrix computing which is easy to implement and makes sense in our context.

Somewhat more elaborately, we use the clients to pre-compute sporting events, for example, chariot races modeled with accurate 2D physics. Think of SETI@Home: we push a data blat to the client with the stats of the chariot teams, their starting positions and race strategies; the client computes the race move-by-move, and sends a computed blat back to the server. The server caches these as scripts for eventual playback. To players the races appear to be real-time, but, they’re actually pre-determined by these client-side calculations. The arithmetic really isn’t that complex, but, it’s easy to distribute, and we want to take advantage of whatever server-side processor savings we can.

There’s no inconvenience to players. Consumer computers spend bajillions of cycles idling between keystrokes. We borrow some. Nobody notices.

The major downside is that TriadCity players are locked to a specialized client: for example they can’t use dumb telnet clients. This is an obstacle for blind players. Telnet clients are simple for screen readers to manage. Our GUI clients have been challenging to integrate with screen readers. As a workaround we’ve written a specialized talking client - which many sighted players turn out to prefer! - I use it most of the time. But this isn’t really where we want to live. We could enable telnet clients and forgo NLP for their users. We dislike that idea a lot - it would violate our commitment to enabling blind players as first-class citizens in our universe - but we may go ahead and do it anyway based on player requests. Haven’t decided yet.

Pushing AI to the client keeps the server small-ish. It’s currently an 8-core, 16gb one-unit pizza box in a co-lo, which we intend to migrate to AWS in a few months. Despite all the AI, it can handle tens of thousands of players no sweat. If we actually had that many players we’d be thrilled.

Please bring your friends.

TriadCity uses multiple AI architectures to model human behavior in different contexts.  Even the simplest is pretty sophisticated: the subsumption architecture for in-game automata does a great job of keeping the City vibrant, where AI-driven characters have homes to go to, jobs to be employed at, hobbies to cultivate, and relationships to explore.  At the higher end of the spectrum, we use external bots which log in to the game world just as human players do, mimicking human decision-making and in a very limited way mimicking human speech.

There’s a trade-off which might not be immediately intuitive between in-game and external AI.  While external AI has access to greater compute resources, in-game AI has access to the database.  Think of movement.  An AI-driven in-game character finds its way around the world by querying a weighted directed graph which the game server keeps in memory.  The model goes like this: in-game AI are residents of the city, they’ve been there a while, and they know their way around.  External AI face a different problem.  If we want to model human behaviors, they can’t simply query the database to ask how to get to point B.  Like humans they must explore the world, building their own mental representation of its geography analogous to the paper or Visio maps drawn by players.

Here’s a practical problem external AI characters must solve to survive.  “My canteen’s empty; I’m thirsty.  Where is that fountain I passed a while back?”  A human will probably use ordinary geospatial reasoning to reckon a reasonable path.  A naive AI can walk back through a history of its moves.  A more sophisticated, human-like AI needs an internal representation of where it’s been, what it’s found there, and how to navigate the most reasonable path back.

I’ve recently enhanced our external bots to build persistent knowledge maps of the world as they explore, using the excellent open-source graph database Neo4j.  This is the same technology the TriadCity game server uses.  The server’s use case is extremely simple.  At boot, once the game world is loaded into memory, a Neo4j graph of all of the Rooms and their relationships is made, with properties describing path attributes.  At runtime, AI characters simply query the graph for a path and off they go.  The database isn’t persisted, it’s re-created at each boot, and is updated on the fly as World Builders add or delete Rooms.  The external AI use case is more nuanced, where the bot must build a persisted graph room-by-room as it explores.  Fortunately the Neo4j code is again very simple.

Each time the bot enters a Room, it asks the graph if that Room is already present.  If no, it creates the new Node, populating its properties with information it may want later: is there a water source here, a food source, a shop, a skill teacher?  Along with the new Node, it adds a Relationship from the originating Node to the new one.  A single simple LEADS_TO RelationshipType is all that’s needed.  If yes - the Room was already present in the graph - it adds the Relationship linking the origin and destination Nodes.  With a little work we can add the cost of the move to the new Relationship: cache the bot’s energy before the move and compare it to when the move is complete; make that cost a property of the Relationship.  The robot now has a comprehensive map of all the Rooms it’s visited, the paths between them, and their costs.  Finding the closest Room with a water source is a simple shortest-weighted-path traversal.

Movement behavior is now more intelligent.  Bots already knew not to re-enter the Room they previously left.  Now they can prioritize directions they haven’t yet explored.  If a Room has Exits north, south and west, and the bot’s previously explored north and west, next time it’ll go south.  The resulting behavior is less random, more like the way a systematic human would do things.  Also, the bot can take the least expensive path back to a water source; that is, it’ll get onto the trail rather than blundering through thickets.  Nice.

Survival skills are improved.  The bot now knows the comparative distances to multiple water and food sources, and can compare their relative merits.  It keeps a catalog of shops with their Item types and can return to buy something useful to it once it has the necessary scratch.  It retains the locations of hidden treasures it can revisit if its current strategy is wealth acquisition, and it knows where the ATMs are so it can deposit its loot.  It can potentially assist human players with directions, or bring them food or water if they’re without - I’m thinking of implementing an Alpine Rescue Dog who can bring supplies to stranded players and guide them out of mazes.  Yes, it’ll have a barrel around its neck.  The barrel was previously doable - now the rescue is, too.

There are some nuances to manage.  One obvious one is that in TriadCity, Room Relationships aren’t 100% reciprocal.  Certain mazes are especially problematic, where if you enter a Room from the East, returning to the East doesn’t necessarily take you back to the Room you originated from.  Sneaky us.  So there’s an impedance mismatch, where Neo4j Relationships are reciprocal by definition, but TriadCity reality isn’t always so.  Also, of course, the mazes change.  The solution we’ve implemented for now is to simply not persist maze graphs, although that’s not particularly satisfying.  More thought required.  A perhaps less obvious problem is closed doors.  If the bot finds an exit West, and tries to move through it, but the door is closed, we have to parse the failure message - “Perhaps you should open the <thing> first?” - and delete the Relationship.  The ugly problem is that “Perhaps you should open the <thing> first” can apply to either doors or containers: “Perhaps you should open the refrigerator first?”.  We can try to disambiguate by keeping a cache of the last several commands, doing our best to figure out whether a move succeeded or not.  But Threads are our enemy here, where events and messages are pushed to us from the server, and the Thread driving bot behaviors doesn’t stop for ‘net lag.  The danger is that the graph could become garbled if the bot fails to recognize that while it intended to move, it in fact did not.  We address this today by periodically walking the bot to a known location; if that traversal fails to arrive at the expected Room, we delete the portions of the graph that were added since the most recent successful walk-back, and begin graphing again from wherever we happen to be, with the hope that before long we’ll meet up with Rooms we already know, and can link the new graph segment to the old.  Not entirely pretty, and we’re open to suggestion.

None of that is Neo4j’s fault.  The graph db is a stellar addition to our external AI.  It lets the robots “think” similarly to the ways humans do.

Here’s how one of our bots thinks of Sanctuary Island, the exact geographical center of the City.  Experienced TriadCity players will note that its north and south poles are reversed.  But that doesn’t matter at all from the bot’s POV.  The point is that it can now easily return to the Temple of the King to level, or the Fountain to refill a canteen, and so on.  This image was generated by Neo4j’s built-in graph visualizer, which is pretty spiffy:

Embedding Neo4j this way was again extremely simple.  One class, a few dozen lines of code, easy-peasy.  My one complaint is that Cypher, the Neo4j query language, has been a moving target.  The change which tripped me up is related to revised indexing schemes which first deprecated, then eliminated Cypher constructs such as “START”.  The first edition of O’Reilly’s Graph Databases is out of date; the one I’ve linked to should be current.  These syntax changes cost me some Internet research.  Time to buy the new edition.

I’ll close by noting an interesting possible future enhancement, now that we have Neo4j embedded with the robot code.  Seems the perfect platform for implementing decision trees.  The bots I’ve been describing are rules-based, via Drools.  Enhancing their decision-making realism via decision trees is attractive - I’ve long planned to explore this idea for modelling addiction in-game.  Perhaps our external bots will be the prototype platform for this type of AI.  Watch this space.

[Addendum 2016.05.21: view of Sanctuary Island and part of the Northwest Third from space.  Doesn’t this graph look like a three-dimensional sphere - a universe?]

Serverless computing is the obvious endpoint of contemporary technology evolution.  From onsite hardware to self-managed datacenters to virtualization to public cloud, the trend resolves in a very attractive way with the end result that your business logic executes “somewhere”, you don’t need to worry where.  An automated platform manages scalability, availability, and security; you’re billed only for the cycles you use, not for the idle headroom you once needed to provide.  On AWS the future is Lambda: we’ve chosen database reporting for our initial experience.

In the old days we’d run a suite of weekly cron jobs on the database server firing sql*plus scripts, piping their output to mailx.  Business analysts, product owners and technology managers would receive human-readable columnar reports in their inboxes, which they’d study and archive however they liked.  Pretty straightforward.  It’s possible to duplicate the same concept in the cloud, and this is initially how we started: install mailx on one of the MongoDB secondaries, have cron run weekly jobs, mailing not particularly human-readable JSON output to the usual suspects - thought being we’d get that JSON human-formatted real soon now.

This strategy takes poor advantage of the AWS platform.  A more sophisticated approach is to save report outputs to a weekly bucket on S3, then mail a link to that bucket to the interested parties.  This centralizes the report archive in a pre-organized repository, and gets the weekly email down from a mini-storm to a single message with links.  Easy change, then: instead of piping report output to mailx, pipe it to a super-simple shell script which reads it from stdin, figures out if a new S3 bucket is needed, creates one if yes, and writes a JSON file to the bucket.  Here’s our version, for your Googling pleasure:

#!/bin/bash

# Writes input from standard in to an S3 bucket named by date. # The file name is passed as param $1 to the script. # 2016.04.04 MP.

DATE=`date +%Y-%m-%d` S3_BUCKET=<base_directory>/$DATE/

# Create the bucket if it doesn't exist: if aws s3 ls "s3://$S3_BUCKET" 2>&1 | grep -q 'NoSuchBucket' then aws s3 mb s3://<base_directory>/$DATE/ fi

# Write stdin to a new file in the bucket: aws s3 cp - s3://sm-reports/$DATE/$1.$DATE.json --content-type text/plain --grants read=uri=http://acs.amazonaws.com/groups/global/AllUsers

Now we simply add a concluding cron job to send an email with the links.  Easy-peasy.  No need for mailx anymore: let’s use AWS SES instead, which offers enhanced security - no more mailx credentials inside a pseudo-user’s home directory.  Instead, SES access is managed by IAM.  Nice.

This outcome already leverages serverlessness.  S3 is an abstracted service, where users pay no attention to concepts of physical location, disk capacity, backup, redundancy, scale, or Web server infrastructure.  SES and IAM are also serverless.  They’re “somewhere” but we don’t care where.  That’s a start.

But here’s the thing.  As noted, the reports are flippin’ ugly.  MongoDB’s JSON output is serial rather than columnar, and there’s no built-in utility analogous to sql*plus which can easily convert JSON blobs to spiffy tables.  So let’s fix this.  We’ll write our own down and dirty little formatting tool converting JSON arrays to tabular HTML.  Where should that tool run?  Lambda’s perfect.  With Lambda we don’t maintain or pay for EC2 instances with inevitable idle time, we simply let Lambda fire up our converter as the weekly cron jobs require.  It’s an event-driven compute pipeline, where cron fires reports, pipes them to a script which writes them to S3, S3 notifies Lambda that new JSON files are written, Lambda converts them to HTML, and SES mails a notice.

I chose to write the converter as a Node.js script.  My inexperienced guess was that this would be less cumbersome than a Java program, avoiding the hassles of frequently uploading new iterations during development and generally, as the lore insists, being faster to work with.  I was wrong in every assumption.

First, non-trivial Node.js programs can’t be entered directly into the Lambda inline editor because necessary Node.js packages aren’t available there.  In our case, the “async” package is required to manage the asynchronous programming model of AWS API calls to S3.  So, you know: le sigh.  I ended up spending about 12 hours on the converter, where some of that time was my Node.js learning curve, but far too much of it was the inevitable save-zip-upload-test-logscan cycle which even Node.js non-noobs are likely to experience.

Second, lack of compiler support made the cycle inevitably more onerous, as every little typo or missing semicolon made it all the way through the loop, to bubble up in the log rather than under my fingers while typing the code.  While I realize that Node.js has considerable real-world momentum, I can’t imagine using it for nontrivial production applications.  I strongly suspect that any large-scale production code base can be written in Java in one-fourth the time, and will be easier and less expensive to maintain over its life cycle.  Experience will tell, of course.  But you heard it here first.

Additionally, there’s a little bit of lag, often 15-20 seconds, between running a test iteration and viewing Lambda’s log output in CloudWatch.  For retrospectively scanning production logs, this will seldom be a problem, and in fact CloudWatch logs themselves can be processed by Lambda functions, a pretty great thing.  During development it’s a hindrance.

Summary: the experience was more of a chore than it should have been.

Ultimately, the manual save-zip-upload-test-logscan process can’t scale to production.  The AWS CLI includes APIs for pushing functions to Lambda.  Developers who prefer the command-line can homebrew their own Continuous Integration / Continuous Delivery scripts.  Larger shops will want to automate their CI/CD pipelines with Jenkins or equivalent.  This’ll be prominent in our enhancements list.

Future TODOs: add a step generating a single HTML index file with links to the reports; mail a link to that index file, instead of two dozen links to the individual reports.  Spiffy up the report format, which right now is down-and-dirty unstyled HTML circa 1994.  Get the Lambda functions into Git and add the CI/CD pipeline to Jenkins.  Consolidate multiple report formatters to a single converter, presumably relying on metadata such as column names added to the scripts generating JSON report output.  You could reasonably sum these up with the single word “professionalize”.  My POC is pretty amateur.

But it works, and that’s impressive.  No production infrastructure to maintain, just pure code execution which automatically scales, is inherently secure, and is guaranteed always available.  For the few cents a month which runtime will cost, this is really seriously something.  Lambda can be used in the event-driven architecture described here, or can be fronted by AWS’s API Gateway service providing RESTful front-ends to Lambda functions.  Perhaps the most serious downside to the Lambda ecology today is lack of persistent database connections: every db request requires connection/authentication/authorization, which is likely too much lag for many real-world production scenarios.  You have to suspect that’ll be fixed soon.

Very psyched to enter the new world.

-----------------------

Addendum 2016.04.21:

In retrospect, it would have been simpler to design a one-step solution from the outset.  For example, a Python script to query the db, transform JSON to HTML, and write the final result to S3.  If I were starting from zip, that would be attractive.  But I started with MongoDB “native” JavaScript queries, so, with those already in place, it was simpler to leverage S3′s event-driven integration to Lambda for the HTML transformations.

Since the original post I’ve consolidated what began as individual functions per report to a genericized formatter accepting directives passed as input; it’s now the responsibility of the JavaScript query scripts to pass those configuration variables to the HTML processor.  This makes it easy to add new reports - no need to update the HTML converter.  I’ve also found a package called “node-lambda” (https://www.npmjs.com/package/node-lambda) which promises to eliminate or at least simplify the inefficient code-save-zip-upload-test-logscan loop I naively began with.  Report to follow if it works out.

But lastly, I’m pretty sold on abandoning Node.js for Python.  Node’s asynchronous model is just frankly cumbersome.  The resulting code is ugly, hard to scan, error-prone, and slow to work with.  The Python is also quirky, particularly pymongo’s departures from canonical MongoDB shell syntax - “find_one” instead of “findOne”, and different quoting rules.  But it’s clean, and straightforward to work with.  Watch this space.

SmartMonsters’ Web site was one of the first in our experience to be fully dynamic.  Every page is a JSP.  Written c. 2001, the goal was similar to far more elaborate turnkey systems such as Kana: content personalization, particularly ad selection, based on user profiles. That model lost interest for us as advertising became unviable on smaller Web properties.  Nevertheless it still has practical uses.  With minimal effort we can reconfigure displayed pages for users who self-identify as visually-impaired, removing all images, simplifying navigation, and replacing sighted TriadCity clients with those designed to be blind-friendly.  That’s really not at all a bad thing to be able to easily do. Architecturally, the site suffers from age-related drawbacks we’d like to fix while migrating.  It’s not designed for Cloud: it doesn’t scale out in an elastic way.  Sessions are old-school server-side JEE requiring replication between instances.  Single-instance content caches minimize database roundtrips, but ensure that instances will be out of sync with one another.  The forum software in particular caches aggressively, guaranteeing both read and write inconsistency between instances.   Fortunately these drawbacks are straightforward to address. Sessions: Elastic Beanstalk will sync ‘em for us, invisibly to our apps, using DynamoDB as back-end store.  Sold: we’ll just turn that on and be done.  But note, we had to put some thought into this option, as it would have been equally feasible to use memcached - see below.  Memcached is tried-and-true, straightforward to configure, and well-understood in practice; in theory it might be a little faster.  I opted for DynamoDB because that’s the solution which AWS explicitly supports.  They provide the necessary JAR, it’s simple and correctly documented.  I don’t think performance will suffer to any degree that users could perceive: DynamoDB’s really, really fast.  So I went with the supported option.  There’s one detail which bugs me, though: the DynamoDB version doesn’t seem to clean up after itself.  There are thousands of rows of old sessions in the table.  I’m not excited re having to write my own script to prune these.  Perhaps it’ll do it itself after enough time elapses.  So far no-one on the AWS fora have answered my question about this. Caches: in-memory caches which previously used Ehcache can be migrated to memcached on ElastiCache.  Elastic Beanstalk server instances can then share a central ElastiCache cloud, restoring sync.  Why memcached over Redis?  Because the cached Forum data are frequently Maps or serialized objects.  No, I didn’t design this.  The main Java Redis client, Jedis, doesn’t store Objects; the main memcached client, Spymemcached, does.  That settles that.  The good news is that our architects have thoughtfully exposed the caching API via the Strategy OO design pattern: there’s a CacheEngine Interface with multiple concrete implementations.  All I had to do was write a MemcachedCacheEngine, confining our migration effort to this one happy location.  Not too bad. There’s an ancillary upside to having memcached in the mix: offloading cache memory to its own cloud allows the main app servers to downscale.  We pushed the cache memory elsewhere, so the app servers don’t need as much.  This makes it inexpensive to generalize caching and sharing of dynamic information which is costly to compute.  For example, we display a page of aggregate and weekly stats from the TriadCity universe here.  These data originate from weekly Elastic Map Reduce jobs which write their results to the primary MongoDB; from there they’re pulled when this page is displayed.  We can eliminate the MongoDB roundtrip by caching them in memcached after the first request.  Stuff like that.  These are very minor performance tweaks and they do add the complexity of additional moving parts.  But, since we have those parts now anyway, why not take advantage? These trivial migration issues were easily solvable.  If we stop there, though, we haven't fully taken advantage of the architectural best practices AWS calls for.  IMO, AWS excels at decomposing and distributing application logic among multiple components communicating via message queues.  Analogously to offloading cache to secondary instances, a lot of the processing which a beefy app server would formerly have done can be distributed to small, specialized environments.  Especially when these jobs can be initiated asynchronously, there’s a net benefit of overall downscaling: the app servers can be smaller when they don’t require headroom to process blocking tasks while remaining responsive to new user requests.  Elastic Beanstalk exposes this architectural pattern via the concepts of “Web tier” and “worker tier” instances, which it’ll link for you with SQS queues automagically.  The Website app offers a couple of opportunities for this kind of decomposition: First, we can offload routine email notifications to a worker tier.  New membership confirmations, password changes, and so on: these can be dropped fire-and-forget onto a message queue from which a minimally-powered worker instance can process them asynchronously.  These messages are infrequent, so the worker instance can be very tiny.  I sized it as a T1 Micro and it’s perfectly happy. [Update: after a mystery production hang traced to an out of memory error, I upped the instance to S2 Small.] On a larger scale, we can extend this strategy to email newsletters.  AWS offers a Simple Email Service handling exactly these kinds of bulk blasts.  Coolness.  In this case, the instance should be larger, but, we don’t need it to be permanently online.  We only need it while we’re broadcasting newsletters, which is at most once a month.  The design can center on another queue-based background job, passing newsletter content to a worker tier and letting that call SES.  We’ll leave the worker instance stopped most of the time, waking it up only when needed.  Ideally this should all be scripted, so that the Servlet which listens for newsletters to be input can fire up the broadcaster on demand - and shut it down again when the blast is complete.  Not super high priority, so we’ll do this manually for the time being.  Cool idea, though. Finally, let’s serve static content from CloudFront.  Images, primarily.  We’ll upload them little by little, converting the Website app into a traditional Web 1.0 three-tier kinda thing, with CloudFront doing what Akamai or another static content service would historically have done.  Again, a very minor performance tweak, and again, adding minor complexity through extra moving parts.   Here’s what the complexity buys, in my view.  The primary Web servers can be smaller.  The work is now distributed among a cluster of scaled-down components interoperating in our VPC.  Although we now use triple the number of virtual machines, overall monthly cost is about a third less than the previously more vertical design because the individual components are dialed way down.  By distributing and scaling out we create a more nimble system at lower operating cost; and we can scale the components independently.  IMO the minor additional operating complexity is worth the trade. There were two unexpected architectural gotchas:   First, the TriadCity game server exposes its list of currently-logged in players via RMI.  The Web site JSPs call into the game’s RMI server for the dope.  Reasonable when all the moving parts are co-located.  Not a happy first choice when those parts are separated by the Internet.  Really kinda ugly when communicating through NAT.  After a day of mucking I pulled the plug.  Current logins and TriadCity date can be grabbed from the database, so that’s that.  Server version, boot time and other details can be written to DB when the game server boots, and the JSPs can pull those data from there too.  An hour, ish.  Done. Second, turned out we saved some interesting non-Serializables to HttpSession.  Loggers, and some ugly Freemarker utilities courtesy yet again of those who designed the open-source Forum software we’ve heavily adapted.  Loggers, NP: just grab them statically when needed, don’t save a class instance.  Freemarker: blecch.  Half a day detangling the mess the Forum designers bequeathed us in their zeal to insist that their design not be embeddable.  Well - it wasn’t their thing.  Fixed after struggle. One AWS quirk caused some temporary grief.  CloudWatch monitors CPU and network I/O, but not memory or disk use.  Apparently these are not natively available to their hypervisor.  Our Web server farm in Elastic Beanstalk went unstable for part of a day: sometimes our frequent new deployments would succeed, other times fail, with little indication in the various AWS logs aside from cryptic return codes.  After much head-scratching I decided to run yum update just for a brain break, which surprisingly led to the solution: out of disk.  Launch new AMIs with double disk allocations; terminate the old instances; back to happy.  Monitoring strategies will have to be figured out for disk and memory use - these are more important to our application contexts than CPU or I/O. I bravely tested elasticity with simulated traffic against the production environment.  Worked as advertised: additional instances spin up automatically, and sessions are shared.  Nice!  Then I terminated a production instance - manual emulation of Chaos Monkey.  EB brought a replacement up within a few seconds.  Very nice.  This lets us easily enable another AWS best practice: keep EC2 instances young.  AWS says that odds of instance failure grow with an instance’s uptime.  The implication is: terminate and restart instances frequently.  With Elastic Beanstalk and shared sessions we can do this completely transparently to users.  I see no reason not to write a script which terminates an instance every day.  If your testing allows confidence that EB will replace the lost instance invisibly to the world, why not?  Let chaos be your friend.

This migration was happily straightforward.  Most of the work centered on simple decompositions taking advantage of AWS best practices.  There was no downtime at all.

Nice!

For years we’ve pushed Web content from desktop to production without human intervention: a very early example of what’s nowadays known as “Continuous Deployment”.  A simple cron job on the Web server updated from CVS every fifteen minutes.  And that was it: if you checked-in new JSPs, they were gonna go live, baby!  It was up to you to test them first.  Code as opposed to content was always released the old fashioned way: 100% manual. We can do allkindsa better. The Cloud-based process we’ve implemented begins with traditional Continuous Integration, but goes beyond it to fully automate Continuous Deployment of every component where that’s a reasonable goal.  The TriadCity game server is the only part of our universe we wouldn’t want to auto-deploy - we wouldn’t want to bump players off every time we push code.  All of our other services and environmental components, from Web sites to robots to SOA-based AI resources are now deployed with full automation without downtime every time we push code.   Here’s the flow:

Developers are responsible for comprehensive tests.  Test suites must pass locally before developers push to SCM.  This remains a manual step, but as you may have personally experienced, those “You broke the build!” emails are good incentive to follow the process.

Push to SCM triggers a Continuous Integration build by Jenkins in The Cloud.  Post-commit hooks in SCM call back to Jenkins, which pulls, builds, and tests.  Build failures or test failures trigger nastygrams and iteration.

If the tests pass, Web sites, robots, AIML interpreters and other services are immediately deployed by Jenkins directly to production.  Most of these are Elastic Beanstalk instances: EB will do its upgrade-and-swap dance for us invisibly.  Jenkins’ Elastic Beanstalk Plugin simply writes the new artifacts to shared storage on S3, then calls Elastic Beanstalk’s REST API to trigger the swap.  EB handles the roll-over invisibly to end-users.  Very cool.  Note that there was no real work in this for us: we simply configured the plugin to deploy to EB as a post-build step, and that’s that.

For some products, acceptance tests follow deployment.  That’s backward - we’ll fix it.  For example, the AIML files instructing many of our chatterbots are validated in two steps, first for syntactic correctness during the standard test phase, second for semantic correctness after deployment, when it’s possible to ask the now-live bots for known responses.  We do this on live bots purely because it's simple: we’ll need a separate acceptance test environment to make it correct.  There'll be an intermediate step: on build and unit test success, Jenkins promotes into the acceptance test environment where it runs the final semantic checks.  Only on success there will Jenkins finally promote to production.  Failure anywhere stops the promotion train and broadcasts nastygrams.  We’ll run a handful of final post-deploy tests as smoke tests.  This is a compelling vision I’ll make real shortly.

Re the TriadCity game server, as noted we don’t want to deploy to production automatically because it would bump users off at awkward and unannounced moments.  TC scales up, not out: it’s a single large server requiring downtime for code boots.  We’d rather trigger these manually.  Here, we have Jenkins remote-trigger a shell script on the TC server which calls mvn to pull the latest build from Artifactory in The Cloud.  That build sits on the TC server and it’s up to a human to bounce the game.  Because we pull-build-test-and-deliver following every code push, there may be many builds delivered without ever being launched.  TC’s administrators will launch when they choose to.  The good news for them is, the latest build’s already present on their server; and it’s guaranteed to have passed its tests.  Just run the bounce script and you’re good to go.  You can reasonably call this “one-button deployment”, which is still a pretty great thing.

We can perform one-button rollbacks on any of the Elastic Beanstalk instances.  Simply re-deploy an older build via EB’s Web console, and that’s it.  The sole requirement for safe rollbacks is to never make backward-incompatible schema changes in the databases.  Since most of our DBs are Mongos - schemaless - this is easy-peasy.  We can similarly roll-back the TriadCity game server with a single script. Elastic Beanstalk impresses me a ton.  Watching it “rolling update" to a new version of the Web site is pretty hecka gratifying.  The spinner goes gray for a minute and when it’s green, users arrive at the upgraded site.  None of us touches a thing.  Just push code and the world changes.  Never any downtime.  Invisible to users. SmartMonsters has nothing resembling traditional release dates, not even agile release trains.  We release continually, often many times a day.  At the end of our working day we stop our Cloud instances serving Jenkins, Artifactory and other resources: no need to pay for what we’re not using.  It’s like turning the lights out when you leave the room. I believe this is the future.  With 100% server virtualization it’s now possible to bundle the environment with the release, or to put it another way, for releases to include both code and platform.  This vision is far more radical than that of the DevOps movement.  I believe it’s inevitable.  This is not quite exactly how we do it today - instead we let EB provision new environments for us.  The result is identical.  New code running on new virtual instances, many times a day.

TriadCity’s robots Death, Help, Marvin and Oscar are AIML-based chatterbots architected monolithically, meaning each bot is an individual application, fully vertically integrated: TriadCity client code, AIML interpreter, AIML files, plus each bot’s thin layer of hardwired individual behaviors, packaged independently of each other, deployed separately, where each deployment has its own directory hierarchy, and so on. This is not some massive scalability bottleneck.  We’re only talking four bots.  With a little work we can even semi-automatically update their AIML without having to fully rebuild each application. But, there are advantages to decoupling their AIML brains from their TC client code:

 Simpler to develop a new generation of bots hosted on different platforms, where these new bots can access a central shared AIML brain via different protocols and devices.  For example, we can access Oscar’s newly RESTful brain from mobile devices, Web sites, Twitter feeds, and the existing TriadCity bot without having to maintain separate AIML repositories for each client.  AIML updates become independent of the Java code for these multiple platforms, doable in real time without rebooting the bots.  The AIML service can scale independently, becoming fully elastic in The Cloud.

Easier to harden our AIML implementation.  Today we rely on each developer to get it right.  This is a problem, because the XML parser is brittle.  If an AIML topic is rejected, subsequent loading from that file stops: all the following topics are ignored.  This means that from a file containing 15,000 topics, only the first few hundred might be loaded if there’s a simple XML typo like forgetting a closing angle bracket.  Also, notice to developers is non-automated: someone has to read the logs.  If by contrast we deploy the AIML interpreter as its own service, we can use standard CI to validate new AIML centrally, independently of developers, in multiple passes.  Like this: check in new AIML —> Jenkins fires a build —> the AIML is validated with the same SAX parser the interpreter uses at runtime —> if the validations fail, the build stops and developers are notified; if the validations succeed, a remote script triggers a pull from SCM, updating the runtime AIML —> the app is redeployed, invisibly to end users, without downtime —> a final round of functional tests written in the SoapUI tool validates the AIML rules at runtime, catching any regressions.  This is really pretty nifty.

Easier to swap-out or enhance AIML interpreters without having to rebuild and redistribute an army of dependents.  Specifically, we’re unhappy with the “Program AB” reference implementation, which is brittle and poorly performant.  We’re likely to either replace it or fully rework it soon-ish.  Decoupling that engine from its clients helps to future-proof.

Enables the AIML engine to become fully “elastic” in The Cloud: scaling out or in automatically under load.  Nice!

Here are the migration steps:

Pull the AIML interpreter out of the bot code, into a standalone application accessed as a RESTful service running in The Cloud.  Very super minorly tricky, because the off-the-shelf AIML interpreters for Java are file-centric, reading from and writing to directories requiring access.  Granted we’re not interested in re-writing the interpreter to be more flexible, this rules out Google App Engine, which has no concept of file system.  So, Elastic Beanstalk.  With a little bit of hoop-jumping we can make that work.  A shell script in the ec2-user’s home directory pulls revised AIML from SCM; Jenkins will fire that shell script remotely.  Alternatively, we could package the interpreter code and AIML files together in a single WAR which we'd redeploy with every change.  Although this is simpler, I decided against it since our AIML developers are used to accessing their source files in a separate location in SCM.  It may be necessary to revisit this, as updates to EB now rely on knowing the underlying EC2 instances’ DNS addresses, which of course can change elastically under load.

Redesign the AIML interpreter to serve all of the bots.  Our legacy vertical monoliths were each-bot-integrates-its-own-interpreter.  Here, we’ll serve all AIML from one central RESTful service on Elastic Beanstalk.  The REST API will be super-simple: a GET call parameterized with the bot name and the input String.  Return an easy-peasy JSON object.  That’s really it.  Whip it together in JAX-RS and there it is.  Let EB scale it if that ever becomes necessary.

Remove the AIML interpreter from the Web site, having the appropriate JSPs now RESTfully GET their conversations from the service on EB.

Remove the AIML interpreter from the TriadCIty bots, having them similarly call into the service.  I decided while I was at it to consolidate the four existing bots into a single JAR rather than continuing to build each as a separate project.  This is simpler, and there’s very little Java code distinguishing the bots, really just some hard-wiring of default responses and some variation of the circumstances in which they’ll intervene in chats.  A configuration file determines which bot to run; the consolidated code figures out which bot-centric classes to load.

Configure Jenkins to manage Continuous Delivery.  The CI build fires when updates are made to the central SCM.  The build lifecycle now includes two cycles of AIML validation.  A first pass ensures XML correctness before changes are pushed to production; a second pass validates the AIML rules by comparing expected to actual outputs.  If we were a larger company I’d break CD into a deployment pipeline where we'd push first to a functional test environment to run these SoapUI-based tests, acting as a quality gate to production. I may circle back and do this at a later time.  For now we’ll just do it in production, and send ourselves an email when a test fails.

Now let's add some PaaS sugar.  We can log inputs which fail to match any AIML rules, triggering default responses.  This feedback will let our botmasters know what their users are interested in, and how their bots are performing.  Since we’re on AWS, let’s log to DynamoDB, with a super simple schema: bot name, <input>, <topic>, <that>.  Then let’s schedule a job to report on these failures, say, weekly.  Coolness!  Let’s additionally modularize and abstract this logging so that we’re not 100% married to AWS; we’ll do that with the standard OO Strategy Pattern, via an Interface with a Factory delegating to concrete implementations servicing different environments.  Later if we want to bolt from AWS we can simply implement a new Strategy class for wherever we go without untangling miles of upstream code.  Check. The migration took two working days, following the steps above.  AWS basics are straightforward, with intuitive controls.  To my surprise, DynamoDB logging required a full third day.  I struggled to get connected - turns out the published Developer Guide is missing one key point, and incorrect on another.  To connect you have to specify the AWS region, duh.  And, Integer keys are not necessary, Strings are fine, and there's a handy "@DynamoDBAutoGeneratedKey" annotation in the Java SDK which'll auto-gen the key for you.  Nice.  Pleased to get it going: I'd very much prefer NoSQL solutions, especially when the data are inherently non-relational, as these records unavoidably are.

One more area of enhancement seems obvious.  Writing these log messages directly to db is synchronous: our RESTful API server will block during the transaction.  Let's put a message queue between the API server and the database: we'll fire the log messages and forget them.  Easy-peasy: add a "worker tier" to the application; call the write-to-DynamoDB logic from its POST method; implement a new Strategy object which writes log messages to SQS instead of database.  Half a day. The platform is now in place upon which we’ll build a new generation of bots more advanced than these.  There’s an upcoming project to rebuild our QA bots via the Drools rules engine - QA bots seem a natural venue for rules-based AI and with Drools we can invent a Domain Specific Language for QA, enabling simple-ish future expansion.  Let’s get on it. The next Cloud migration step will be SmartMonsters’ fully-dynamic Web site.  More to follow.

Although SmartMonsters does not require a large server environment, it makes sense to move much of it to The Cloud.  As we know, the obvious advantages are reliability, availability, and elasticity.  Until recently these benefits were outweighed by cost.  Our awesome colo provider, Osiris Communications, have been able to offer VPS services which are far more cost-effective than comparable IaaS options from AWS; while the beefy TriadCity server, still hosted on our own hardware, would be cost-prohibitive to virtualize.  Increasingly, though, we believe the cost equation has shifted far enough in favor of AWS that we’re slowly migrating our Web infrastructure there. Despite the relatively small scale of our needs, migration is by no means a trivial project.  To take proper advantage of, for example, Elastic Beanstalk, our aging and by today’s standards monolithic Web app needs to be broken into decoupled components which can scale independently.  Here’s one thorny example.  The Web site’s forum section is powered by a forked and heavily modified open source code base now deeply integrated into - tightly coupled with - the rest of the site’s SSO, authorization and persistence frameworks.  The problem is that this third party code aggressively caches to local memory. Multiple instances would not sync state, so user visits from one click to the next would likely differ.  It’s inappropriate to deploy this forum code on EB.  It has to be pulled out of the Web application and deployed independently as a single instance. Or a more trivial but still useful example.  Our clever chatterbots Oscar, Marvin, Help, Death and Eliza all currently run in small EC2 instances, logging in to the TriadCity server in the same way that humans would.  We can make them “elastic” by creating a group of one for each bot.  If it were to die, AWS would make a new one for us.  That’s fine, but because the bots are monolithic we’re unable to share their infrastructure in a way that lets us easily add new apps which reuse their code.  This is wasteful and cumbersome.  Instead, we could break the bots into components, for example by exposing via EB a shared AIML interpreter with RESTful interface.  Then, the Web site, the bots in TriadCity, and perhaps a new generation of bots running on mobile devices could all call into a shared and elastically scalable AIML service. This problem is shared by many of the companies I’ve encountered in my consulting practice.  Years of legacy architecture, often very large scale, designed to be “integrated” back in the day when that was a marketing bullet rather than a swear word.  Or, evolved into “integration” over years of unplanned enhancement.  One successful company I’ve recently spoken with intends to solve this issue by building a new generation of monoliths running in AWS.  Yes, it’ll be The Cloud.  And, it’ll be a super-hip new language beloved by 25-year-olds. But it’ll be the same problems transferred to new context. So, SmartMonsters’ migration will simultaneously be a major re-architecting of the Web site application, breaking it down wherever possible into loosely coupled components communicating through REST.  At the same time, the TriadCity bots and other network-based utilities such as SCM, email, database servers and so on will move out of colo onto Cloud-based services.  I’ll write several posts here outlining how it goes. Initial State Here’s what we start with:

SCM is by CVS - remember that? - on a VPS in the colo.  This SCM manages not only source code but also company docs.  Deployment of new or edited JSPs to the production web server is automated by CVS update via cron.

Email is self-administered Postfix, on a VPS in the colo.

MongoDB instances are spread among VPS and physical hardware in the colo.  There’s a primary and two secondaries.

PostgreSQL is on a VPS in the colo.

A single Tomcat instance serves the entire Web site, which is 100% dynamic - every page is a JSP.  This instance experiences a nagging memory leak, probably related to the forum code, which no-one has bothered to track down yet; we simply reboot the site every night.

The TriadCity game server is physical hardware, with 16 cores and 32gb RAM, racked in the colo.

The Help, Oscar, Marvin, Death and Eliza chatterbots run on small EC2 IaaS instances on AWS.

Jira and Greenhopper are on-demand, hosted by Atlassian.

DNS is administered by the colo provider.

Firewall rules, NAT, and port forwarding are also administered by the provider.

Anticipated End State Here’s where we think we’re going:

SCM by Git, hosted by a commercial provider such as Beanstalk, GitHub or similar.

Company docs moved off of SCM, onto Google Drive.

Email by Google, maintaining the smartmonsters.com domain and addresses.

MongoDB instances hosted in EC2, geographically dispersed.  There’ll continue to be a primary and two secondaries.  Backups will be via daily snapshots of one or another secondary.  It may be optimal to occasionally snapshot the primary as well: Amazon recommends shorter-lived instances whenever that’s feasible.

PostgresSQL via RDS.

The Web site forum software broken into its own EC2 instance, accessed via RESTful interfaces called by the parent Website app.

The remainder of the dynamic Web site served by Elastic Beanstalk; sessions shared via DynamoDB.  At least two instances, geographically dispersed.

Static web content such as images hosted on S3 and served by CloudFront.

Long-term backups such as legacy SQL dumps stored on Glacier.

The chatterbots broken into two layers: an AIML interpreter, probably an adaptation of Program AB, accessed via REST by the Java-based bot code which is other respects is very similar to any other TriadCity client.  The AIML interpreter will probably be served by Elastic Beanstalk; the bots can continue to run from EC2 instances.  The interpreter will have to be adapted to save state in DynamoDB; the AIML files can live in S3; but note that the program will have to be adapted to also be able to read files from the local file system, in case we choose to deploy it in a different way.  We may or may not make those instances elastic.

Jira and Greenhopper remain at Atlassian.

A micro EC2 instance will serve an Artifactory repo to be used by Continuous Deployment.

DNS will probably move to Route 53; we’ll first discuss this with our colo provider.  Advantage to us is self-administration.

The TriadCity game server will remain on physical hardware inside the colo; it’s simply not yet cost-effective for a company as small as SmartMonsters to virtualize servers of that scale.

Firewall rules, NAT and port forwarding for the TriadCity server remain with the colo provider.  Security groups, key management and so on for the AWS infrastructure will be self-administered via IAM.

Monitoring of the AWS resources will be via CloudWatch.

The AWS resources will probably be defined by CloudFormation, but, we don’t know enough about it yet.

Continuous Deployment will have to be re-thought.  Today, a shell script on the Web server runs CVS update periodically, a super-simple CD solution which has worked with no intervention for years.  The virtualized infrastructure may allow more sophistication without much additional cost.  For example it may be sensible to add Jenkins for CI, then automate deployments from there which not only push code, but create new Beanstalk instances while terminating the old ones, a strategy which will keep the instances always relatively young per AWS best practices.  Beanstalk will keep these deployments invisible to end users, with no app downtime.  More exploration is in order.

As I write this, we’ve already begun with the databases and some of the low-hanging fruit.  I’ll write about those in the next installment. Conclusion This is a journey which many companies, small and large, will be taking in the coming years.  The advantages of resilience and automated elasticity are simply too strong.  With costs as low as they’ve become, I believe the default for most enterprises will be some form of Cloud-based deployment.  I hope that these migration notes turn out to be helpful to you.

The relational paradigm maps poorly to OO design.  Objects don’t have relations.  When you use an ORM for CRUD you end up with a denormalized schema lacking the core point of *relational* databases.  If your OO design relies on inheritance you can potentially end up with absurdly large numbers of tables.  In TriadCity we have hundreds of Item subtypes with varying properties; mapping these to traditional normalized RDBMS would result in spaghetti schema.  Somebody has to win: you end up with either a lousy database or lousy OO design.  Because MongoDB is schemaless these problems don’t arise.  You can put all those hundreds of Item subtypes into the same MongoDB Collection, no worries.  And you can report on them in a perfectly granular way,  Schemalessness removes the impedance mismatch between the two views of the world. Our migration turned out to be a far larger project than my initial tests led me to expect.  It was trivial to migrate game object persistence as noted above.  Without hyperbole, an hour or two.  But of course, we do vastly more with data than merely game CRUD.  For example, we're beginning to use rich data analytics of 14 years of in-game auctions to enable AI driven characters' bidding behavior: when to bid, how much, what ceiling.  Previously that logic would have lived in a vendor-specific stored procedure, and would have been much less comprehensive at runtime than what we can achieve now; today instead it’s a combination of MongoDB plus Neo4J for real-time lookups.  I’m happy as can be to have this logic live in Java inside the game engine, rather than a proprietary stored procedure language in the database.  But, implementation in the new technology took some effort. I chose not to merely replicate our previously relational schema in MongoDB.  It would have been incrementally easier to simply map tables to Collections 1-for-1 via some down-and-dirty little Java snippet, and that would have been that.  Instead, I chose to try to take advantage of MongoDB’s JSON-based ability to store related information in a single record via nested arrays.  Here’s an example.  Our RDBMs used normalized tables to relate basic user account information to: demographics; settings and preferences; security roles and access restrictions; a listing of bounced email addresses; and a table of locked accounts.  So, a parent table and five children, related by the unique ID of the parent account.  Perfectly reasonable in a SQL world.  In MongoDB, it makes better sense to keep all of this related information in a single record.  So our MongoDB JSON Person record has an array field for demographics, another for roles and privileges, another for settings and preferences, and simple boolean flags for bounced emails or global account lock.  Easy-peasy.  This consolidation is straightforward to code to, and appealingly enough it results in a human-readable document easily browsable in your favorite administrative GUI.  Nice.

Almost as a side effect, the need for transactional guarantees simply goes away. In a SQL world most transactions are second-order: they’re really artifacts of the relational technology itself, where an update to a Person per above might require writes to multiple tables. Post-migration we literally have not a single use case where a transaction is necessary. We do rely on write and replication guarantees in MongoDB: our writing Java thread blocks until it receives an appropriate write acknowledgement. Usually we ask for a write-plus-replicate acknowledgement. I suppose that’s sort of the MongoDB analogy to relational transactions but they’re conceptually different and simpler in practice. The Java driver for MongoDB handles them automagically.

Object-oriented queries in MongoDB's Java API turn out to be super simple to internalize.  They read like almost-English and they make perfect sense.  Here’s an example and its SQL equivalent. This looks up the most recent sale price of a Player’s in-game house.  MongoDB:

    DBObject where = new BasicDBObject("playerHouseID", playerHouseID);     DBObject fields = new BasicDBObject("price", true).append("timestamp", true);     DBObject sort = new BasicDBObject("timestamp", DESCENDING);     DBObject found = this.mongo.db.getCollectionFromString("PlayerHouseSales").findOne(where, fields, sort);     if (found == null) { // this player house has never been sold         return -1;     }     return (Long) found.get("price");

The SQL equivalent minus whatever framework boilerplate you employ would have been something like:

    "SELECT * FROM (SELECT price FROM player_house_sales WHERE player_house_id = ? ORDER BY timestamp DESC) WHERE rownum = 1"

Personally, I find the OO version vastly easier to conceptualize. Object-oriented queries like these make it straightforward to eliminate the 200 or so stored procedures which had tied us to one specific RDBMS vendor.  Nearly all of these stored procedures existed because it’s easier to implement branching logic in a stored procedure language than in SQL.  Typically the branching test was whether a row already existed, thus insert or update.  The MongoDB Java driver’s save() method handles upsert for you, automagically converting between insert and update as the case warrants.  A few circumstances which are marginally more complicated are easily handled in Java.  I experience this as a huge gain.  All of our data persistence and analytics logic are now in Java in a single location in a single code base - no longer scattered among multiple languages and disparate scripts in different locations in SCM.  Nice. I don’t know what the best practice is for sequences.  Sequences aren’t recommended for _id unique keys.  But, there are circumstances where an incrementing recordNo is handy.  For example, I like to assign a recordNo to each new Room, Exit, Item and so on.  A dynamic count() won’t do for this, because these objects are occasionally deleted after creation, e.g., a world author changes her mind about a Room or series of Rooms.  I’d nevertheless like to know the grand total of how many Rooms have ever been created, without keeping a separate audit trail.  It seems straightforward to create a sequence record in the database with one field for each sequence; you then call the $inc operator to do what it sounds like.  Since this is not the _id unique key we're talking about, I ran with it.

I'd like to see more syntactic sugar in the query API.  Common operations such as summing or averaging columns are cumbersome today.  There's a new-ish Aggregation Framework seemingly designed for large-scale offline reporting ala your favorite SQL report writer.  For ad hoc in-game queries, I'd prefer the Java API, and would like to have common operations like summing or finding max or min values highly runtime optimized, as they are in the RDBMs we're replacing.  There are corners of the MongoDB programming world which to me seem immature, like these.  Development seems very active, so hopefully these gaps will be filled soonish.

I've encountered one problem not yet solved.  Sparse unique indexes don't seem to work.  Inserting more than one record with a null value into a field with a sparse unique index fails with a duplicate value error on the null, which the "sparse" part is supposed to finesse.  Google suggests others have experienced similar glitches.  For the time being I've turned off the index, and am continuing to research.  Really do want this working. Runtime performance is excellent.  Superior in most cases to the SQL equivalent.  For example, loading 18k Rooms from RDBMS including their default Items and Mobiles typically required about 40 seconds; with MongoDB it’s usually 15 seconds or thereabouts.  On my MacBook Pro with 8 cores and 16gb RAM, a complete game boot in the SQL world took around 60 seconds; it’s half that with MongoDB.

[For a client, I’ve subsequently done a shootout of very high-volume write processing between MongoDB and Oracle, where “high-volume” means hundreds of millions of writes. MongoDB was an order of magnitude faster under serious load.]

MongoDB’s system requirements are more demanding, though.  MongoDB likes to put the entire data set into RAM.  It assumes a 3-node replica cluster with several gigs memory per node.  Our data set really isn’t that big, so we’ve got it shoehorned into t2.small instances in three regions on AWS.  Without a lot of hands-on lore, I can’t say yet whether we’ll turn out to need to bump up these instance sizes, or if they’ll be fine.  Film at eleven. [Update 2014.10.31: instance sizes unchanged.  All's well. Your mileage will vary.] MongoDB’s automated replication is pretty nifty for high-availability and for offloading queries from the writeable system.  We run our analytics queries using map/reduce against mongodumps taken from one of the secondaries, freeing the primary for production writes.  Nothing’s necessary to enable this behavior, it works out of the box.  For very large data sets, sharding is easy-peasy.  There's a handy connector linking MongoDB with Hadoop, for our offline map/reduce queries.  The database is designed for these topologies, which in many RDBMS worlds are cumbersome to achieve. Now for the interesting future.  Granted our enhanced ability to easily analyze our 14 years of comprehensive log data, we can begin to enable vastly more powerful and realistic in-game AI.  Here’s an example.  The data show us which Items sell most frequently at auction; what the most successful starting prices are; what the average, minimum and maximum final auction prices are for each Item type; how many bids are common; upper and lower bounds on all these things; medians; distributions; curves; etc. etc.  They can help us predict which Players or AI-driven characters might be interested in buying which things.  By pulling the “big data” from MongoDB and graphing it with Neo4j, we can enable super-zippy real-time queries like, “Should I bid on the current auction Item?  If yes, how much?  What should my top bid be based on the bid history so far?  If I buy it at the right price, will I later be able to profit from re-auctioning it?  When’s the best time to re-auction?  Are there any Players currently online who may be interested in the Items in my possession?  Are there any robots or other AIs who might be interested?”  This has far-reaching implications not only for the quality of in-game AI, but also for our ability to expand and make the in-game economy more rich and realistic.  Along with in-game merchants who buy Items from Players, we can now have AI-driven characters buy and sell from auction, creating another sales channel from which Payers can benefit.  Sweet.

A note re Mongo and JSON. If your company models its business objects in JSON, Mongo or another highly-performant JSON database makes great sense. It’s far more straightforward than shoehorning JSON into a relational model that was designed for an earlier world, and belongs there.

SmartMonsters now is almost-but-not-quite entirely post-SQL. We’re using “polyglot persistence”, as the hip term has it: the principals are MongoDB, Neo4J, DynamoDB, with our legacy third-party forum software still trapped on PostgreSQL. Very much looking forward to a world where AWS offers a managed MongoDB service which we don’t have to provision and monitor for ourselves.

TriadCity’s sophisticated “subsumption architecture” for NPC behaviors depends on accurate, comprehensive knowledge of game world geography.  Unlike traditional MUDs where NPC movement is statically confined to home zones and is largely random, TriadCity is alive with highly individualized NPCs capable of purposeful navigation throughout the City. For years the code enabling this directed movement has been a simple homegrown weighted graph mapping all possible paths from every point A to every other point B.  Weights were based primarily on movement costs by alignment.  Separate graphs were computed for Evil, Good and Neutral paths, plus specialized graphs for Cops and for residents of the Capitoline neighborhood - two categories of NPCs with restricted movement possibilities.  The graphs were computed at server boot, following instantiation of all Rooms and the Exits linking them.  They were static: unchanging after initial computation, essentially freezing NPC knowledge of world geography at the state of its existence at boot.  And they were monolithic: an NPC could look up a path from the Neutral graph, or the Cop graph, but not a hybrid NeutralCop graph.  For a long time this was “good enough for now”.  While inflexible, it enabled accurate if not always subtle NPC navigation from any arbitrary origin to any chosen destination.  NPCs simply looked up their path and, following Dijkstra, off they went. TriadCity eventually outgrew this simple design.  We want to be able to accurately model arbitrary restrictions on NPC paths, for example, excluding NPCs from certain streets based on Role, or gender, or other classification.  It's important for NPCs to react to more than one category of restriction.  And we want their movement behaviors to be as “human” as possible, so, not necessarily always taking the shortest possible path, if the shortcut for instance leads through someone’s private garden, or in one pub door and out another.  Usually real people would go around the garden, or stay on the sidewalk outside the pub.  We grew especially frustrated with our simple system's inability to track dynamic changes to the game world at runtime.  It would have required a major project to extend our down & dirty proprietary code to handle these features. Instead, we’ve chosen the excellent NoSQL graph database Neo4j as our solution.  While the approach is essentially the same - a weighted graph based on Dijkstra - Neo4j’s comprehensive query-ability allows us enormous flexibility at runtime.  It easily enables the kinds of ambitions described above, with impressively little code. We run Neo4j in embedded mode.  Easy-peasy, since our server’s written in Java.  The database is populated during server boot after Rooms and Exits are loaded.  Nodes are Rooms, Relationships are built from outbound Exits leading to destination Rooms.  The algorithm’s straightforward.  Iterate the Rooms, define a Node for each; iterate a second time, catalog each Room's outbound Exits, look up the destination Room each Exit leads to, find the Node representing that destination, and define a Relationship between the origin and destination Nodes.  We only need one RelationshipType: LEADS_TO.  As each Relationship is defined we flag it with boolean properties representing the restrictions we want to impose: isNoCops, isNoBeasts, isNoHumans, isNoDeathSuckers, isPerformerOnly, isCapitoline, hasDoor, and so on; these properties belong to Exits and destination Rooms, so we simply look for them and set the flags appropriately.  At runtime, NPC path lookups are via a PathFinder built by passing GraphAlgoFactory’s static dijkstra() method a custom CostEvaluator which looks for the boolean property flags defined on Relationships and totals costs accordingly.  For example if the calling NPC has the Cop Role, and a Relationship is flagged isNoCops, the CostEvaluator will assign a prohibitive cost to that Relationship - we arbitrarily use 100,000 - effectively removing it from consideration.  When PathFinder searches for the least-cost path it’ll choose one which bypasses that expensive Relationship.  Managing the scenario where we want NPCs to go around a pub although the shortest path leads through it is straightforward: we simply assign a cost of 100 to all Relationships flagged hasDoor, so that generally NPCs will prefer not to move through doors unless they’re the only paths available.  Nice. Implementing Neo4j required just a single class of about 300 lines, and took about half a working day including time spent tuning the costs. Runtime performance is excellent.  The Neo4j solution takes longer to bootstrap - about 5 seconds for 18,000 Rooms on my MacBook Pro, compared to just a few milliseconds for our deprecated home baked version.  But, its runtime lookup performance is superior.  We judge overall AI processing load by tracking average completion times of our one-per-second Behaviors pulse which triggers Behaviors throughout the game world.  Average Behavior computation on my laptop was steady in the range of 120 to 150 milliseconds with the old grapher; with Neo4j it’s been consistently around 80, an efficiency gain of 50% or better.  Memory use looks about the same. So far, runtime movement appears to be perfect.  Everybody goes where they’re supposed to, and they now do it with the subtle enhancements highlighted above.  Importantly for us, it’s now very easy to tune movement by simply adding new property flags to Relationships and looking for them in our custom CostEvaluator.  And, we can easily enhance the system to adapt to runtime changes in the game world itself, for example when builders add new Rooms or 4d mazes rearrange themselves.  This now sits high on the priority list. I’m impressed with how easy it’s been to include Neo4j in our project.  The API is intuitive and very nicely documented.  I was able to take and tweak a Dijkstra example from the Neo4j web site with very little effort.  Very pleased! Along with Neo4j, we’re currently migrating away from SQL to MongoDB and Hadoop for game object persistence and analytics.  Watch this space!

All NPC behaviors in TriadCity are event-driven.  Not just NPC AI, but all change everywhere in the virtual world, from the motion of subway trains to room descriptions morphing at night, is event-driven.  Here's a quick sketch of some TC event basics. In TriadCity every action is encapsulated in an Event object.  This is straightforward in OO Land.  Take for example player-generated actions.  Player Commands are expressed in the server code as Strategy Pattern objects encapsulating the necessary algorithm.  Each Command will generate one or more Events as its outcome(s).  For example, the Steal command generates a ThiefEvent, which it parameterizes differently depending on success or failure.  Commands instantiate the appropriate Event type at the appropriate moment within their algorithms, handing off the new Event instance to the current Room, which propagates it wherever it needs to go.  Read on. Here's the basic structure:

public Interface Event {         public long getID();         public long getEventType();         public long getVisibility();         public String getSourceID();         public String getTargetID();         public String getRoomID();         public String getMessageForSource();         public String getMessageForTarget();         public String getMessageForRoom();         .... }

The Steal command will instantiate an Event type called ThiefEvent, populating it with the information shown above and a ton more.  The eventType will vary depending whether the Steal succeeds or fails; the assigned visibility will be computed depending on the Thief's Skill level and multiple variables.  The messageForSource will be: "You steal the silver coins from the drunken Texan", or "You fail to steal the silver coins from the well-dressed man, and are detected."  Appropriate messages are computed for the target and for bystanders in the Room.  When complete, the ThiefEvent is passed to the Room, which in turn passes it to everybody and everything currently present. What happens next? Up to world authors. One form of NPC behavior is triggered when Events are received.  The Behavior instance tests the Event against its triggering conditions.  Am I supposed to fire when the Event is a ThiefEvent and its getEventType() == STEAL_FAILS_EVENT_TYPE and the NPC I'm assigned to is the Event.getTarget()?  Or in English, did someone just try to steal from my NPC and fail?  If the world author has assigned the TriggeredAttackBehavior to an NPC and given it those trigger conditions, then the Thief will be attacked when those conditions are met.  This is a form of rule-based AI, expressed in OO patterns. Who'll see what happened?  Depends on their SeeSkill and other attributes.  If observers are skilled enough, they may be able to watch Thieves stealing from victims.   The victim won't see it happen, but observers with high enough SeeSkill might.  One of many many forms of character subjectivity in TriadCity. OO allows TriggeredBehaviors to be niftily flexible.  Behaviors can be triggered if they receive any MovementEvent, or only if the MovementEvent.getEventType() == BOW_EVENT_TYPE.  So: any movement in your Room and you'll jump to your feet on guard.  But, you'll only Bow if someone Bows to you first. All Behaviors are Event-triggered, even ones which appear random, for instance pigeons moving around the Gray Plaza.  Although their movement Behavior is indeed semi-random, it's triggered by a worldwide pulse called the DoBehaviorsEvent, which is propagated everywhere in the game world once per second by an Executor running a background Thread.  Every Behavior tests that DoBehaviorsEvent: do I ignore it, or randomly do something, or do something not randomly?  TimeEvents trigger actions such as closing shops and going home, and so on. Events are integral to Player subjectivity.  One form this takes is that Player objects include a Consciousness Strategy object via composition.  That Consciousness object acts as a filter on the Events passed through it.  It has the ability to modify or ignore the Event.getMessageForXYZ() methods it encounters, or generate entirely new ones.  This is one of the ways we can make you see things which maybe aren't even "there". Events in TriadCity are like the chemical "messages" flying around your brain, telling you when to breathe and fire your heart muscle and decide to go online to play games.

I recently found a discussion of programming languages on a MMORPG hobbyist bulletin board.  Much of the lore there was garbled, revolving around old but persistent myths of the performance characteristics of C, C++, and Java.  I thought I'd share some hard-won wisdom. To first state my claim to bonafides in this realm.  In real life I'm a cloud scale software architect at the senior executive level, meaning I've got a couple of decades' experience building really big commercial systems with bazillions of users.  Plus I've got twenty years hands-on writing TriadCity, which I hope I can say without boasting is the most technologically sophisticated textual RPG to date. So here's the true dope. Under super heavy load - truly Massively multiplayer scale - Java has a slight advantage over other languages in scalability and perceived responsiveness because under load its garbage collection mechanism makes more efficient use of CPU cycles than hand-rolled memory management.  Yes, I realize that contradicts the claims that many developers will make, but, it's not hard to understand.  If you profile an Apache web server under very heavy stress, you'll find it spending a large percentage of its cycles calling malloc() and free(), where every call requires some number of CPU instructions.  At its core it's just doing lots of String processing, but there you go: memory in, memory out, many many millions and millions of times once the load gets going.  At some point, memory management will crowd business logic out of the processor, and perceived degradation is no longer graceful. By contrast, Java allocates its heap in just one go, when its runtime initializes.  Garbage collection is a background process and on multi-core systems does its thing pretty efficiently.  CPU time can be dedicated to business logic instead of system overhead.  This plus team productivity are the two reasons Java dominates the market for enterprise application servers. To stress some qualifiers.  A well-designed and written system in C or C++ will outperform a badly designed one in Java.  "Better" is contextual: do you care about throughput, latency, graceful degradation under stress, or some other criterion?  Maybe you're super proficient in one language over another: use that one. Note in general that OO languages are a better match for game servers than procedural ones.  I started TriadCity in C back in the day because at that time I was more proficient in C, and there was plenty of open source C game code to model things on.  But I found in practice that some of the concepts I wanted to express just didn't "feel" intuitive in C, and it became harder and harder to be productive in the code base once it got large enough.  Switching to Java solved these problems because thinking in objects maps more intuitively to virtual world concepts; and because, at least in my experience, it's far easier to organize a very large OO code base for extensibility and maintainability.  I recently read a different hobbyist RPG board where the lead developer of a very popular old-school MUD written in C complained about the effort required to implement a new character class.  I'm pretty sure I can tell you why it was so rough for him.  I don't think it would have been as much of a chore in TriadCity. To underscore this point: OO code bases scale better.  Today TriadCity has about 40,000 server-side classes.  No, seriously.  But I can find them all easily, because organizing OO code is straightforward, and Java's package concept helps.  More importantly, thoughtful OO design insulates you against brittleness, enabling evolution.  Adding that new character class would have been less daunting in an OO world.

Uniquely to any "game" we're familiar with, TriadCity imposes elaborate subjectivities on its players. This means that if you and I walk our characters into the same Room at the same time, we might see it described differently depending on our characters' attributes, skills, moral alignment, history, and other variables.  The type and degree of subjectivity is largely controlled by the world author who wrote the Room.  Most authors tread lightly, but there's no reason they can't make things wildly different if that makes sense for their context. It also means that one of our characters could be Schizophrenic.  We can impose any manner of "unreal" experiences on characters, from highly distorted Room descriptions to conversations with other characters who don't actually exist. The code which enables these two forms of subjectivity is pretty simple.  Both forms rely on the Strategy OO pattern, applied in different places.  I'll briefly sketch them both. Rooms, Items, Exits, Characters all share common attributes inherited from their parent class, WorldObject.  These include Descriptions, Smells, Sounds, Tastes, and Touches.  I'm capitalizing the nouns because these are Java Interfaces which WorldObject possesses by composition.  A code snip for illustration:

public abstract class WorldObject extends Object() {           protected Descriptions descriptions;           protected Smells smells;           protected Sounds sounds;           protected Tastes tastes;           protected Touches touches;           public final String getDescription(WorldObject observer) {                     return this.descriptions.getDescription(observer);           }           public final String getSmell(WorldObject observer) {                     return this.smells.getSmell(observer);           }            public final String getSound(WorldObject observer) {                     return this.sounds.getSound(observer);           }            public final String getTaste(WorldObject observer) {                     return this.tastes.getTaste(observer);           }            public final String getTouch(WorldObject observer) {                     return this.touches.getTouch(observer);           } }

Descriptions, Smells, Sounds, Tastes and Touches are Strategy Pattern Interfaces defining the simple contract which will be implemented by a rich hierarchy of concrete classes.  At runtime, the correct implementing class is assigned to each WorldObject according to the directions given to that WorldObject by the author who wrote it. For example, a world author may write a Room and choose to have its descriptions vary by hour of day.  Using our authoring tools, she assigns it the TimeBasedDescriptions type, and writes a description for dawn, dusk, day, and night.  At server boot, that Room will be assigned the type of Descriptions she specified, and populated with her writing.  The TimeBasedDescriptions class has the responsibility of figuring out what time it is whenever its getDescription() method is called, and returning the right one. That's not subjective, though.  Although the descriptions vary by time of day, the same description will be shown to every character who enters the Room.  If she wants to, our world author can be more adventurous, assigning concrete Description types from our library.  She could choose AlignmentAndTimeBasedDescriptions which will show descriptions which vary by both time of day and character moral alignment; or Descriptions which vary by Gender, or health condition, or attribute status.  Each concrete implementing class knows how to calculate the description which is appropriate for the observer.  As you can see, the code for WorldObject remains extremely simple. A second and often more radical flavor of subjectivity is implemented by a Strategy Interface called Consciousness.  Each character has a composed Consciousness object which acts as a filter, transforming messages received via game Events, or not, according to its logic.  The TriadCharacter class includes these snips:

public abstract class TriadCharacter extends WorldObject() {           protected Consciousness consciousness;           public boolean proccessEvent(MovementEvent evt) {                     return this.consciousness.processEvent(evt);           }           public boolean proccessEvent(SenseEvent evt) {                     return this.consciousness.processEvent(evt);           }            public boolean proccessEvent(SocialEvent evt) {                     return this.consciousness.processEvent(evt);           } }

Concrete Consciousness implementations include SleepingConsciousness, FightingConsciousness, SchizophrenicConsciousness, IncapacitatedConsciousness, and many others.  Their job is to analyze incoming Events and determine whether to pass along the messages associated with those Events unchanged, or changed.  A simple change might be to simply squelch an Event if the character is asleep.  A radical one might be to change a ChatEvent or TellEvent, or even concoct entirely new ones, if the character is experiencing SchizophrenicConsciousness.  In these cases, the Consciousness type is assigned by the server at runtime, not by world authors, and it'll be swapped out for a different one as circumstances change.  So these transformations are extremely dynamic. Using the Strategy pattern via composition allows the class hierarchy to remain reasonably flat and really very simple.  It would be possible for example to create a massive class tree resulting in things like TriadCharacterWithDescriptionsThatVaryByAlignmentAndTimeWithSchizophrenicConsciousness.  But for obvious reasons that would be a nightmare to maintain.  Composition keeps the WorldObject tree simple; Strategy objects chosen at runtime provide the rich library of behaviors.

I suppose that mazes are a staple of the MUD tradition.  Starting with Adventure, you're underground and you're lost.  Throw in some orcs and you're in Moria.  Whatever. We follow the tradition with lotso mazes, but, we try to make them interesting.  This post'll be about the technology rather than the content, but, we really do try to make the content interesting too. TriadCity Maze types are implemented via the Strategy OO pattern.  There's that word "Strategy" again.  We do a lot of that.  A Java Interface called Maze defines the contract; multiple concrete implementations can be swapped out at runtime.  Among the implications: world authors not only have a ton of maze algorithms at their disposal, but we can actually swap out algorithms at runtime whenever that seems the thing to do. TwoDMaze is a simple x/y grid of Rooms, always in a rectangular pattern.  World authors can specify how many Rooms wide versus tall, and the Builder tools will prompt them if they provide the wrong number of Rooms.  Thus a TwoDMaze defined as x=7 y=5 has gotta have 35 Rooms assigned, no more, no less.  TwoDMaze puts the Rooms in consistent order across the grid, that is, Room 1 will always be the upper left corner, and the Rooms will be assigned in left-to-right order at every server boot.  Exits between Rooms will be assigned dynamically according to what maze-drawing algorithm is chosen by the world author.  You guessed it, these algorithms are again Strategy objects.  We've got Kruksal, Prim, Depth-First, and a pile of others.  Exits will be computed and distributed around the maze at server boot. Add a z dimension and it's a ThreeDMaze.  Have the Exits dynamically reassigned at random or semi-random during server uptime and it's a FourDMaze - think Harry and Cedric racing for the cup. We can vary all of these by not assigning the Rooms in consistent order.  Just shuffle the Room[] before laying out the grid.  Now there'll always be a Room 1, but it won't always be in the upper left corner.  It could be anywhere, meaning that special fountain or other object of interest will have to be searched for anew after every server boot.  Or shuffle them every hour, or shuffle the grid and the Exits and really confuse people. But wait, there's more.  We don't have to always assign ordinary FreeExits - the Exits that allow movement between Rooms without impediment.  They can be RandomDestinationTeleportationExits or any other Exit type we like, in whatever proportion we like.  Yes, there's one extremely notorious maze full of nasty critters with pointy sharp teeth offering only a single teleporter to get out - no way you're going back the way you came.  Bring a large party with supplies.  Find that teleporter before you run out of grenades. More still.  Authors can assign Behaviors to the Rooms or the Exits, making things even more interesting.  TelGar's writing a big slag heap you have to climb to reach something valuable.  The slag heap kicks up tonso dust so there's no telling exactly where you are, and because the ground is loose, you're constantly being forced downward as the earth slides out from under you.  It's a compelling adventure, enabled by combining a dynamically-changing Maze type with a SlideRoomBehavior which forces you to move in a certain direction, like a slide in a playground.  You can eventually get to the top, but it's a struggle. These Strategy Pattern objects allow the server code to remain tidy and easy to maintain, while putting a lot of power into the hands of world authors.  Authors can assign whatever crazy algorithm best suits their thematic purpose.  Just how lost do you want the punters to get?  You've got lots of options.  Somebody invents a new maze-generation algorithm?  Great, we'll add it.  Easy-peasy. Death to mazes like Moria, --Mark

We use a pile of different AI strategies to drive NPC behaviors in TriadCity.  Most of these are quite simple, AI being in my opinion far more artificial than intelligent.  One of our most successful strategies to date is based on Rodney Brooks' "subsumption architecture" which we've borrowed from robotics.  Here's a short description of how we use it in TriadCity. Briefly, subsumption architecture rejects the approach of modeling intelligence via symbolic representation, choosing instead to mimic lifelike behaviors by building high-level behaviors from layered libraries of low-level components.  Lowest-level components might be as simple as "I hit something"; the next layer up might include several avoidance strategies such as "back up and try a slightly random vector", or "move left several inches and try again"; a higher level still might abstract from particulars to "explore the world".  The idea is that behaviors become more abstracted from details the higher they are in the hierarchy, which in the end looks something like a pyramid, with "behave like a human" at the top. In TriadCity we implement subsumption by composing higher-level abstract behaviors from an increasingly vast library of detailed lower-level ones.  Low-level examples are "open the door", "look around the room", "pick up an object".  You'll note that these correspond neatly to TriadCity's library of player-typable commands.  An intermediately abstract behavior might be "move from current location to the Virtual Vegas Casino", which involves querying a directed graph, moving along the retrieved path, opening any doors encountered, and so on.  A higher-level behavior might be "play slot machines in the Virtual Vegas Casino", which would require successfully moving to the correct location in the game world and executing the Playslot command.  Higher still might be "gamble in the Virtual Vegas Casino", which would allow a randomized selection between playing slots or roulette or etc.  World authors don't need to interact with the subsumed low-level concrete behaviors.  They just assign the abstract high-level ones and the architecture manages the details. A common example in TC is the assignment of weighted "behavior groups" based on time of day.  Let's say you assign your NPC two behavior groups, one for the dawn, day, and dusk hours, and a second for night time.  Your daytime group may include sending the NPC to work, or having it hunt slave killers in Zaroff Park, or eat lunch at a favorite restaurant, or watch the chariot races in the Circus Maximus.  You'd assign each of these possibilities a weight determining its likelihood of being chosen, plus a weight determining its likelihood of being swapped out for a different one.  By night you might have your NPC eat dinner somewhere, go see a movie, go shopping, go back to hunting its favorite slave killers, or go home to sleep.  The result is really quite rich.  If you follow this fellow around, he performs his varied activities with a lot of human-like unpredictability.  He won't simply circle the same static course over and over, and so for example if there's some NPC you'd really like to stalk, you'll have to go find him first because he could be all over the place doing things that are appropriate to his character.  This is very easy for world authors to work with, allowing rich individualization of non-player characters without a ton of effort.  It puts a pile of AI sophistication into play without forcing authors to interact with it. These "subsumption" behaviors contribute a good deal of verisimilitude to the game world.  In TC you don't get a lot of NPCs standing around waiting to be killed.  We even send our judges home from court at night.  There's continual motion and variation: you can stand on a popular street corner people-watching NPCs on their way to and from jobs and hobbies and murders and freeing slaves.  This is one of the aspects of the TriadCity project I'm most proud of.

TriadCity's command vocabulary is far larger than the 400 or so commands typed by players.  Every Admin and Builder action is implemented as a command, so there are tens of thousands altogether. In C, you'd probably route user input through a simple switch() statement.  CircleMUD does this.  The parser and command router are tightly coupled inside one master controlling statement which switch()es on the user's first word of input.  Logic can be parceled out to functions, or handled right there inside the switch() statement itself.  This is fine for a couple of hundred commands, but growing it to tens of thousands would be a performance and maintenance disaster. We ensure infinite command routing scalability for TriadCity by combining the Strategy and Hashed Adapter OO patterns. Strategy encapsulates an algorithm as an object, allowing algorithms to be selected and substituted at runtime.  In Java you'll typically implement Strategy via an Interface defining the algorithm's contract.  Classes implementing that Interface define the available algorithms.  In TriadCity we have a Command Interface defining our Strategy contract:

public Interface Command {           public abstract boolean doCommand(<multile-args-go-here>); }

And we have many thousands of classes implementing that Interface, for example, Chat, which implements the doCommand() method to iterate all NPCs and logged-in Players, sending the arguments to the chat channel. The Hashed Adapter pattern achieves scalable routing by stashing these concrete Command instances in a HashMap keyed by command name.  Each concrete Command is a Singleton; we populate the map with command names as keys, instantiated Command Singletons as values.  The resulting routing algorithm is super-simple:

Command command = this.commands.get(commandName); command.doCommand(<multiple-args-passed-here>);

There's other logic wrapped around these lines, for example authorization checking for security. But that's the idea. By using Hashed Adapter this way, routing scales infinitely without performance degradation.  Server processing is very efficient, and for developers there's no nightmare switch() statement of ten million lines to maintain.  By using Strategy, it's very easy to do the file housekeeping necessary to keep tens of thousands of Command implementations easy to navigate.