You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@flink.apache.org by "David (Jira)" <ji...@apache.org> on 2026/05/04 06:48:00 UTC
[jira] [Updated] (FLINK-39599) Expose pendingRecords metrics in FLINK Kinesis Connector
[ https://issues.apache.org/jira/browse/FLINK-39599?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
David updated FLINK-39599:
--------------------------
Description:
The autoscaler (post-FLINK-31326) uses pendingRecords as optional input for catch-up sizing. Kafka and Pulsar sources expose this; Kinesis does not, so catch-up.duration is a no-op for Kinesis jobs.
Proposal: Estimate pendingRecords as millisBehindLatest × moving_avg_records_per_ms per split, summed across the operator. Document explicitly that this is an estimate based on recent throughput, not an authoritative count (the Kinesis API doesn't expose one).
> Expose pendingRecords metrics in FLINK Kinesis Connector
> --------------------------------------------------------
>
> Key: FLINK-39599
> URL: https://issues.apache.org/jira/browse/FLINK-39599
> Project: Flink
> Issue Type: Improvement
> Reporter: David
> Priority: Major
>
> The autoscaler (post-FLINK-31326) uses pendingRecords as optional input for catch-up sizing. Kafka and Pulsar sources expose this; Kinesis does not, so catch-up.duration is a no-op for Kinesis jobs.
> Proposal: Estimate pendingRecords as millisBehindLatest × moving_avg_records_per_ms per split, summed across the operator. Document explicitly that this is an estimate based on recent throughput, not an authoritative count (the Kinesis API doesn't expose one).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)