Closed as not planned
Description
In the high-concurrency write log scenario, we found that hot_threads is always on the translog.sync operation. After a long investigation, I found that the application client often consumes several Kafka topics at the same time to form a bulkrequest and write it to Elasticsearch. Each topic corresponds to an index, resulting in a very small bulk quest to each shard, so the translog sync operation is very frequent .
This problem can be alleviated by setting translog async. In order to quickly locate the problem, I think it is necessary to add stats for bulk per shard, including bulk_shard_sizes, bulk_shard_lantency, and bulk_shard_qps.