#hyperdex

Most of Emin Gün Sirer’s posts end up linked here, as I usually enjoy the way he combines a real-life story with something technical, all that ending with a pitch for HyperDex.

The problem here stemmed from the broken-by-design interface and semantics offered by MongoDB. And the situation would not have been any different if we had used Cassandra or Riak. All of these first-generation NoSQL datastores were early because they are easy to build. When the datastore does not provide any tangible guarantees besides “best effort,” building it is simple. Any masters student in a top school can build an eventually consistent datastore over a weekend, and students in our courses at Cornell routinely do. What they don’t do is go from door to door in the valley, peddling the resulting code as if it could or should be deployed.

Unfortunately in this case, the jump from the real problem, which was caused only by the pure incompetence, to declaring “first-generation NoSQL databases” as being bad and pitching HyperDex’s features is both too quick and incorrect1.

1) ACID guarantees wouldn’t have solved the issue; 2) All 3 NoSQL databases mentioned, actually offer a solution for this particular scenario. ↩

In a post introducing HyperDex backups, Robert Escriva compares the different backup solutions available in Cassandra, MongoDB, and Riak:

Cassandra: Cassandra’s backups are inconsistent, as they are taken at each server independently without coordination. Further, “Restoring from snapshots and incremental backups temporarily causes intensive CPU and I/O activity on the node being restored.”

MongoDB: MongoDB provides two backup strategies. The first strategy copies the data on backup, and re-inserts it on restore. This approach introduces high overhead because it copies the entire data set without opportunity for incremental backup.

The second approach is to use filesystem-provided snapshots to quickly backup the data of a mongod instance. This approach requires operating system support and will produce larger backup sizes.

Riak: Riak backups are inconsistent, as they are taken at each server independently without coordination, and require care when migrating between IP addresses. Further, Riak requires that each server be shut down before backing up LevelDB-powered backends.

How is HyperDex’s new backup described:

The HyperDex backup/restore process is strongly consistent, doesn’t require shutting down servers, and enables incremental backup support. Further, the process is quite efficient; it completes quickly, and does not consume CPU or I/O for extended periods of time.

The caveat is that HyperDex puts the cluster in read-only mode for backing up. That’s loss of availability. Considering both Cassandra and Riak promise is high availability, their choice was clear.

Update: This comment from Emin Gün Sirer makes me wonder if I missed something:

HyperDex quiesces the network, takes a snapshot, resumes. Whole operation takes sub-second latency.

The key point is that the system is online, available while the data copying is taking place.