apache
S3 Multipart Upload Throttled by S3 leading to task failures and lag (for streaming ingestion)
Affected Version
33
Description
#17674 added AWS transfer manager for s3 segment uploads. This is enabled by default and will push segments that are over the threshold size in multiple parts.
In practice we have seen that this config being enabled generally works fine with default configuration. However, there are cases where we are getting throttled by s3.
In instances where we see the problem, we are suspicious that cold s3 partitions (created for new segments) are resulting in early throttle from s3 when Druid starts immediately pushing a multi-part segment. Once throttling occurs, index tasks often (maybe always?) fail due to s3 push failures.
- s3 has a 3.5k write threshold for throttling on a prefix, however it has to scale to that threshold and will throttle sooner while the new prefix is warming up. Since druid is creating new prefixes for every segment, we will always be writing to a cold prefix and at risk of early throttling when a multi-part segment push begins
Example Log snippet: Error: com.amazonaws.SdkClientException: Unable to complete multi-part upload. Individual part upload failed: Please reduce your request rate. (Service: Amazon S3; Status Code: 503; Error Code: SlowDown; Request ID:
@suraj-goel for your awareness since you are original contributor! I'll try to add more context as I can, especially for re-producing the problem and potential configurations to better tolerate throttles