Investigation conducted by @nvanbenschoten shows significant impact on foreground SQL latency when executing backfills; summary here (internal). The solution will be to reduce memory usage to reduce GC (assist) pressure.

• Reduce default size of blocking buffer. Suggested default of 1/8 of current default (125MB).
  • Adjust per_changefeed_limit to 128MiB #84686
• kvevent: implement chunked blocking buffer #86421
• changefeedccl: Make kvevent.Event memory efficient #87718
• changefeedccl: reduce size or # of pointers of event struct #84709
• changefeedccl: investigate initial SQL latency spike on initial scan #86469

Jira issue: CRDB-17745

Epic CRDB-24463