Deeply nested aggregations with large dataset


#1

Hello,

We’re looking to use ES for reporting and analytics in addition to search, however I’ve been hitting some limits when running some deeply nested aggregates with large datasets. I’ve tried various approaches in order to overcome this. Unfortunately, we’re limited to one node at the moment, and I understand that there is a jvm limitation of ~32GB.

I’m working with the latest version 6.1.2, using NEST/Elasticsearch.Net on Windows Server 2012 with 32GB and ~400GB disk space.

The data are financial transactions, essentially fact tables that include potentially a large number of dimensions and metrics and are indexed into Elasticsearch. Our requirement is to be able to generate dynamic reports that allow a user to select potentially up to ~20+ “group by” terms. Each fact table is on it’s own index with 3-4 shards, from ~650+mb - 1GB, 600k+ upwards towards 4Million docs each shard.

I’ve gone through various articles on trying to figure out how to optimize, like using doc values, tune various config settings, etc. and I’ve read that it is highly discouraged to do deeply nested aggregations due to the combinatorial problem.

As suggested here


https://www.elastic.co/guide/en/elasticsearch/reference/current/search-aggregations-bucket-terms-aggregation.html#_filtering_values_with_partitions

I’m currently trying to use the terms aggs partitioning to see if I can “page” through the large datasets. Even if I’m able to aggregate one partition at a time, is my understanding correct that the result would not be sorted across partitions?

I’d like to get a better understanding as to how those with terabytes of data are able to (if indeed they are) run large (aka deeply nested) queries on limited memory using ES. Are other technologies used in addition to Elasticsearch in some intermediary phase? Do I restructure the data? Or is it merely a matter of growing and scaling the cluster? Then, in that case, how/can it really manage to aggregate terabytes of data across nodes?

Any pointers to docs/explanations on the inner workings, and eventually solutions in handling large aggregate queries whose size goes well beyond the JVM limit is greatly appreciated.

(The alternative is that we fall back on using cubes on MS SSAS, but we’re hoping to leverage ES as we’re already using it for search).

Thanks very much.


(Mark Harwood) #2

Correct. Partitions by design are an arbitrary and even subdivision of a key space so are not organised by any values that might be attached to those keys. If you want to sort on the keys themselves then take a look at the composite aggregation and the after parameter. Note that the keys offered by the composite agg are limited to values that can be retrieved from a single doc e.g. date interval or a term. It can't be a value derived from multiple documents like "sum" of transaction values.

Let's assume you're trying to get a sorted list of the bank accounts that have spent the most money in the last year. You may have millions of accounts, if you're using a typical logging config you've likely scattered each of their transactions across different shards and time-based indices. This data configuration is not conducive to doing this sort of analysis on millions of entities because there's too much query-time joining of related data to be done across indices and nodes. This is not necessarily an elasticsearch issue - just physics. You may want to look at shifting some of the compute cost to index-time rather than query time and use something like entity-centric indexes to fuse related data together in a localised form rather than leaving it scattered around your network for queries to deal with.


#3

The data has already been "fused" together at the source, so is indexed in ES as either "account details", "charges", "adjustments", "transactions", for example. At this point, I'm just trying to run aggregate queries against any single index at a time.

I'll take a look at composite aggregation, however I see that it is in beta. Is this planned to stick around long term?

Thanks again in advance.


(Mark Harwood) #4

It's worth reading the background to this labelling for some general advice on status of beta features.


(system) #5

This topic was automatically closed 28 days after the last reply. New replies are no longer allowed.