Skip to content

feat(inference): backfill - #16

Merged
UsatovPavel merged 1 commit into
mainfrom
feat/HACK-183-incident-backfill
Sep 28, 2026
Merged

UsatovPavel merged 1 commit into
mainfrom
feat/HACK-183-incident-backfill

Conversation

@UsatovPavel

Copy link
Copy Markdown
Contributor

Jira

HACK-183

Changes

  • BACKFILL_HOURS (default 0 = off). Once per runtime volume (marker runtime/backfill.done) the cursor is rewound to the first event of the last N hours; the normal stream then re-predicts them in id order, so each channel ends with the prediction of its newest event. Expired ones (ts + horizon < now) are skipped as usual but still update windows.
  • Warmup reaches back window + N hours, so rolling windows are full at the backfill start.
  • Why: inference-incident started fresh on the stand and predicts only channels with events after its start (139 channels after 1.5 h).

Rollout

Backend compose sets BACKFILL_HOURS=30 for inference-incident; the next restart runs it once (~200k events).

Verification

  • test_backfill_predicts_recent_events_once: no prediction without backfill, one with it, none again after the marker.

🤖 Generated with Claude Code

BACKFILL_HOURS rewinds the cursor once per runtime volume so the stream
re-predicts events of the last N hours; windows warm up before that
point. A fresh incident service otherwise covers only channels with
events after its start.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@UsatovPavel
UsatovPavel merged commit 248f8d7 into main Sep 28, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant