Will it fail together with the network?
Let us check the Elasticsearch behavior in context of the CAP theorem.
CAP theorem stays that a distributed system communicating over an asynchronous network can’t provide these three properties at once:
This theorem is often used (and misused), but luckily there are some good clarifications over the Internet (see e.g. 1, 2 or even plain english, yet sexist and overly simplified explanation).
Recently, a colleague of mine has sent me an article where the author tests Elasticsearch through CAP theorem lenses. The article, Call me maybe: Elasticsearch, written by Aphyr is a part of the Jepsen series. Aphyr tests how various databases and distributed systems behave in case of network failures. It is worth adding that the blog posts are very detailed and technical and, moreover, fun to read :-).
Quite frankly, I was terrified. The bottom line is that in case of various network failures Elasticsearch stays available, but inconsistent. I.e. clients do not see all of the acknowledged writes (not to mention updates, etc.).
Since network partitions are unavoidable, then according to CAP theorem, once can only build CP or AP system. Looks like in case of Elasticsearch then AP decision was not conscious. E.g. Shay Banon (the creator of Elasticsearch), says this on its mailing list:
I personally believe that “within the same data center”, network partitions very rarely happen, and when they do, its a small set (many times single) machine that gets “partitioned out of the network”. When a single machine gets disconnected from the network, then thats not going to affect elasticsearch.
The documentation is not very verbose on Elasticsearch CAP properties and various people give different explanations on the mailing list. According to Aphyr tests it sometimes acts as AP (transitive partitions), usually it acts as CP (because of the minimum master nodes settings) and sometimes it just drops some writes, but it is not available (so we have only P :-)).
Some people say that you can use Elasticsearch as a primary data source. Don’t do it. As proven in the linked article it can lost your data.
Luckily, there is solace to us working with Elasticsearch on production systems. First of all, you use it to index your data. The data are consistent and safe in your authoritative database. In case of an inconsistency you just reindex. Secondly, Aphyr’s post made Shay and the Elastisearch folks reconsider their system behaviour and fix (at least some of the problems). The fix is already on trunk (pls see this issue).
I think that Elasticsearch is a great tool for its job (you know, for search). Big thumbs up for Shay and his crew and also for the community. As you can see from the linked issue, a good bug report with a set of tests can improve the project greatly.