Working with transforms at scale

Transforms convert existing Elasticsearch indices into summarized indices, which provide opportunities for new insights and analytics. The search and index operations performed by transforms use standard Elasticsearch features so similar considerations for working with Elasticsearch at scale are often applicable to transforms. If you experience performance issues, start by identifying the bottleneck areas (search, indexing, processing, or storage) then review the relevant considerations in this guide to improve performance. It also helps to understand how transforms work as different considerations apply depending on whether or not your transform is running in continuous mode or in batch.

In this guide, you’ll learn how to:

Prerequisites:

These guildelines assume you have a transform you want to tune, and you’re already familiar with:

The following considerations are not sequential – the numbers help to navigate between the list items; you can take action on one or more of them in any order. Most of the recommendations apply to both continuous and batch transforms. If a list item only applies to one transform type, this exception is highlighted in the description.

The keywords in parenthesis at the end of each recommendation title indicates the bottleneck area that may be improved by following the given recommendation.

In order to optimize transform performance, start by identifying the areas where most work is being done. The Stats interface of the Transforms page in Kibana contains information that covers three main areas: indexing, searching, and processing time (alternatively, you can use the transforms stats API). If, for example, the results show that the highest proportion of time is spent on search, then prioritize efforts on optimizing the search query of the transform. Transforms also has Rally support that makes it possible to run performance checks on transforms configurations if it is required. If you optimized the crucial factors and you still experience performance issues, you may also want to consider improving your hardware.

In a continuous transform, the frequency configuration option sets the interval between checks for changes in the source indices. If changes are detected, then the source data is searched and the changes are applied to the destination index. Depending on your use case, you may wish to reduce the frequency at which changes are applied. By setting frequency to a higher value (maximum is one hour), the workload can be spread over time at the cost of less up-to-date data.

Depending on the size of the destination index, you may consider increasing its shard count. Transforms use one shard by default when creating the destination index. To override the index settings, create the destination index before starting the transform. For more information about how the number of shards affects scalability and resilience, refer to Scalability and resilience

If you have defined a transform source index query, ensure it is as efficient as possible. Use the Search Profiler under Dev Tools in Kibana to get detailed timing information about the execution of individual components in the search request. Alternatively, you can use the Profile. The results give you insight into how search requests are executed at a low level so that you can understand why certain requests are slow, and take steps to improve them.

Transforms execute standard Elasticsearch search requests. There are different ways to write Elasticsearch queries, and some of them are more efficient than others. Consult Tune for search speed to learn more about Elasticsearch performance tuning.

Imagine your continuous transform is configured to group by IP and calculate the sum of bytes_sent. For each checkpoint, a continuous transform detects changes in the source data since the previous checkpoint, identifying the IPs for which new data has been ingested. Then it performs a second search, filtered for this group of IPs, in order to calculate the total bytes_sent. If this second search matches many shards, then this could be resource intensive. Consider limiting the scope that the source index pattern and query will match.

Use an absolute time value as a date range filter in your source query (for example, greater than 2020-01-01T00:00:00) to limit which historical indices are accessed. If you use a relative time value (for example, now-30d) then this date range is re-evaluated at the point of each checkpoint execution.

There is no one-size-fits-all sharding strategy. A strategy that works in one environment may not scale in another. A good sharding strategy must account for your infrastructure, use case, and performance expectations.

Too few shards may mean that the benefits of distributing the workload cannot be realised; however too many shards may impact your cluster health. To learn more about sizing your shards, read this guide.

The max_page_search_size transform configuration option defines the number of buckets that are returned for each search request. The default value is 500. If you increase this value, you get better throughput at the cost of higher latency and memory usage.

The ideal value of this parameter is highly dependent on your use case. If your transform executes memory-intensive aggregations – for example, cardinality or percentiles – then increasing max_page_search_size requires more available memory. If memory limits are exceeded, a circuit breaker exception occurs.

Runtime fields and scripted fields are not indexed fields; their values are only extracted or computed at search time. While these fields provide flexibility in how you access your data, they increase performance costs at search time. If transform performance using runtime fields or scripted fields is a concern, you may wish to consider using indexed fields instead. For performance reasons, we do not recommend using a runtime field as the time field that synchronizes a continuous transform.

If you use more than one group_by field in your transform, then the order of the fields in conjunction with the use of Index Sorting may improve runtime.

Index sorting enables you to store documents on disk in a specific order which can improve query efficiency. The ideal sorting logic depends on your use case, but the rule of thumb may be to sort the fields in descending order (high to low cardinality) starting with the time-based fields. Then put the time-based components first in the group_by if you have any, and then apply the same order to your group_by fields as configured for index sorting. Index sorting can be defined only once at index creation. If you don’t already have index sorting on the index that you want to use as a source, consider reindexing it to a new, sorted index.

The _source field contains the original JSON document body that was passed at index time. The _source field itself is not indexed (and thus is not searchable), but it is still stored in the index and incurs a storage overhead. Consider disabling _source to save storage space if you have a large destination index. Disabling _source is only possible during index creation.