Remote segment backpressure

Introduced 2.10

Remote segment backpressure is a shard-level rejection mechanism that dynamically rejects indexing requests when the remote segment store falls behind the local committed segments on the primary shard. With remote segment backpressure, you can prevent the lag between the remote store and the local primary store. The lag can be caused by slow or failed remote store interaction, remote store throttling, long garbage collection pauses, or high CPU utilization.

Thresholds

Remote segment backpressure is activated if any of the following thresholds is breached:

Consecutive failure: The backpressure is activated if there are N or more consecutive failures. The value of N is configurable in remote_store.segment.pressure.consecutive_failures.limit.
Bytes lag: The bytes lag is calculated by adding the sizes of all the files that are present in local committed segments but not in the remote store. Backpressure is activated if the bytes lag is greater than K multiplied by the moving average of the size, in bytes, of the files uploaded after each refresh. The variance factor K is configurable in remote_store.segment.pressure.bytes_lag.variance_factor. The moving window size is configurable through the remote_store.moving_average_window_size setting.
Time lag: The time lag is calculated by comparing the timestamps of the most recent local refresh and the most recent remote store segment upload. Backpressure is activated if the time lag is greater than K multiplied by the moving average of the time taken to upload new segments and metadata files after each refresh. The variance factor K is configurable through the remote_store.segment.pressure.time_lag.variance_factor setting. The moving window size is configurable through the remote_store.moving_average_window_size setting.

Handling segment merges

At every segment merge, a corresponding refresh is initiated. Because this refresh has new merged segments, the bytes lag instantly spikes. To compensate for this spike, the bytes lag and time lag are evaluated only if the remote store is behind the local primary store by more than one refresh. However, backpressure induced by consecutive failures activates regardless of refresh lag (the number of refreshes by which the remote store is lagging behind the local store).