get-result-summary uses a "mega" query to reduce elapsed time for generating a result summary. The is just many, many queries in a single http request, which increases parallelism significantly on opensearch. Compared to individual queries, this is much faster (like 100x). An alternative is async individual queries, but the overhead of the request per single query is far less efficient than a single http request with many queries.
The problem is that we don't limit the number of queries to bundle in the http request, so very large runs (dozens of iterations and hundreds of samples) make either http or opensearch break.
This work will break up the single multi query into more reasonable chunks.
atheurer
commented
1 month ago
get-result-summary uses a "mega" query to reduce elapsed time for generating a result summary. The is just many, many queries in a single http request, which increases parallelism significantly on opensearch. Compared to individual queries, this is much faster (like 100x). An alternative is async individual queries, but the overhead of the request per single query is far less efficient than a single http request with many queries.
The problem is that we don't limit the number of queries to bundle in the http request, so very large runs (dozens of iterations and hundreds of samples) make either http or opensearch break.
This work will break up the single multi query into more reasonable chunks.