Multiple data directories ->parallel search of shards on same instance?

Hi Everyone,

I understand from the ES docs that having multiple shards from the same index on the same ES instance can slow queries down because the 2..n shards from the same index on the same ES instance will be queried serially by ES.

Question--if the shards from the same index are spread amongst different data directories as opposed to being located within the same directory--will this parallelize querying of shards from the same index on the same ES instance?

Please confirm, thanks!


Elasticsearch is perfectly capable of parallelizing queries to multiple shards on the same node. Do you have a link to documentation that says otherwise?
Naturally, each of the shards will potentially be competing for use of the same resources, so, for example, if the node was deployed on a host with a single CPU core, then each participating shard would have to wait for access. The same principle applies to storage - the more the OS provides for parallelization of access to the resources, the more shard activity will be able to exploit that. So, for example, just having multiple data paths on the same mount and file system wouldn't be such a benefit as actually having multiple devices with dedicated controllers and having each of them available as a data path.

Thanks Buster! I don't have links to official documentation, but here's
where I am getting the idea that > 1 shard for the same index on the same
ES node means the shards are searched serially:

with the following key paragraph:

Finally, it should be noted that any more than 1 shard per node will
introduce I/O considerations. Since each shard must be indexed and queried
individually, a node with 2 or more shards would require 2 or more separate
I/O operations, which can't be run at the same time. If you have SSDs on
your nodes then the actual cost of this can be reduced, since all the I/O
happens much quicker. Still, it's something to be aware of.

And there is this writeup @ with
the following key paragraph:

As discussed earlier, in order to run a complete query, Elasticsearch must
run the query on each shard individually. In a one-shard-per-node setup,
all those queries can run in parallel, because there's only one shard on
each node. With multiple shards on the node, the queries for those shards
have to be run serially. The effect of this can be mitigated if the hard
drive the shards are stored on is relatively fast like a 15,000 RPM
enterprise-class drive or, even better, an SSD. Still, it's something to be
aware of.

These two posts appear to indicate that having more than one shard for a
given index on a given ES node will result in the shards being searched
serially. If this is actually not the case, definitely a cool thing, just
trying to confirm so I can setup my cluster correctly.



And here's another post: with the following key sentence:

We do, however, suggest that you continue to picture the ideal scenario
as being one shard per index, per node.

Again, just trying to ensure I setup my cluster per best practices.



TL;DR - improving your storage I/O bandwidth will result in improved search performance, and I'm not specifically encouraging you to have multiple shards per index per node.

Again, this is just flatly false. Multiple shards of the same index (or different indices, it matters not) that are located on the same node can receive search requests concurrently and fulfill those requests independently to the extent that the OS can make the physical resources available to perform them (CPU cores and storage read capacity).
So while it is true that the more shards on a node are managing tasks at the same time, the more likely it is they will experience contention for those resources, the quoted statement is either accidently poorly phrased or just demonstrates a lack of understanding.

The first cited reference states the situation much more accurately, alluding to the contention for disk reads. What's omitted is the fact that each shard has memory-based activity that can be performed before requiring any read, and with sufficient CPUs, those things can be done in parallel by shards. Now, the results from all participating shards must be reduced to a single response, so there's some fuzzy area about the tradeoff between the incremental gain in (say) searching 10 shards of 1GB each vs. 1 shard of 10GB and the reduce overhead difference, but that same comparison is relevant if the shards are on 10 separate nodes or on one node with really great compute and storage resources.


Realistically, you need to think about your total request load (indexing and searching) when considering sharding.
If you expect to have 100 searches per second on that index, now, with 10 shards on the same node, that's 1000 shard searches per second contending for resources instead of 100 shard searches per second, and that's one reason that the prevailing advice is to have fewer shards. Then, to the extent that you are willing and able to allocate additional servers, you can reduce search latency by increasing the number of replicas per shard. If you have 4 of those nodes instead of 1, each with its own copy of of each shard, the load per node would become (roughly, there's some randomization) 25 shard searches per second.

What absolutely never works, though, is when someone with a "fixed" architecture is unhappy with their search performance, and they read it on the internet that more replicas means faster searches, so they just add a replica per shard. Why doesn't this work? Because they already designed their cluster to be the minimum amount of resource that will support their indexing rate and total storage, and still basically respond to searches, so all of their nodes are pretty fully saturated with activity. Adding another copy of every shard just serves to bring their disks closer to being full, increased their heap usage (and garbage collection), increase the amount of indexing load on every node (every copy of the shard actually performs the same indexing that the primary node does), and provide another shard copy available to be searched. But if all three of the shards are on overburdened nodes, it doesn't really matter which node handles the request.

All of that said, though, back to your original question, having multiple data paths defined, to the extent that they aren't really going to contend for the same device, can be a way to reduce the disk read contention that can happen with a lot of concurrent searches on a node.

Please take a look at The Definitive Guide for a very useful overview of the search process. One conclusion you should be able to draw quite quickly is that the search of multiple shards on the same node doesn't behave differently depending on whether those shards are part of the same index.

I hope this helps.


This helps ALOT--thanks so much for taking the time to explain all of this so well, and yes, I will dive into the definitive guide ASAP.

Thanks again for your time!