v20251016-220858

[binaries] metadata lambda to delete existing metadata (#7359)

Delete previous metadata when new wheel is uploaded

This way, if the lambda fails, there will not be stale data

Testing:
checked on something that doesn't have a metadata file already
checked on something that does have metadata file already


I think it might be better if the metadata is uploaded with the binary,
maybe add some s3 metadata to the binary and whl to show theyre in sync,
then the lambda deletes metadata whose sync tag doesn't match the
binary's

If the binary is uploaded manually, then it there won't be metadata, but
I'm not sure how often this happens. Also domains that do custom stuff?
but idk how often this happens either. Also additional packages that get
uploaded to the index?

v20251015-211348

[autorevert] Extract test retries as separate events (#7327)

This helps filter out signals from flaky tests (that succeed after
rerun) and avoid reverting based on them. Such tests are excluded from
HUD, so this makes autorevert more consistent with HUD.

Changes:
- Extract individual test reruns as separate Signal events
- Dedup retains both outcomes
- For each attempt, emit at most one FAILURE and one SUCCESS event
- Skipped tests are excluded from "success"


Extraction result before:

[2025-10-08T18-48-39.575589-00-00.html](https://github.com/user-attachments/files/22786564/2025-10-08T18-48-39.575589-00-00.html)


Extraction result after:

[2025-10-08T21-30-45.676074-00-00.html](https://github.com/user-attachments/files/22786565/2025-10-08T21-30-45.676074-00-00.html)

notice flaky signals like:

pull:inductor/test_compile_subprocess.py::test_remove_noop_slice_scatter_cpu
pull:inductor/test_compile_subprocess.py::test_remove_noop_slice1_cpu

v20251014-205528

[AUTOREVERT] Ignore issues with ci:sev on pytorch for disabling autor…

v20251010-213316

Increase data fetch range for compiler inductor (#7334)

the compiler inductor runs benchmark from 2 times per day to 1 time per
day.
Increase the time range for regression notification in case we run into
insufficient data situation

see change https://github.com/pytorch/pytorch/pull/165185

v20251010-192708

[log classifier] Match against newly added string for consistent fail…

v20251010-102706

[autorevert] Prevent revert requests for open PRs, fallback to notifi…

v20251009-022528

Sync the regression policy with the regression notification (#7301)

v20251006-145405

[AUTOREVERT] OOPS, test code leaking to production (#7300)

Signed-off-by: Jean Schmidt <[email protected]>

v20251006-141907

[autorevert] Fix state logging on revert (#7295)

fixes:
```
[ERROR] AttributeError: 'AutorevertPattern' object has no attribute 'job_base_name'
Traceback (most recent call last):
  File "/var/task/pytorch_auto_revert/__main__.py", line 311, in main
    autorevert_v2(
  File "/var/task/pytorch_auto_revert/testers/autorevert_v2.py", line 84, in autorevert_v2
    state_json = RunStateLogger().insert_state(ctx=run_ctx, pairs=pairs)
  File "/var/task/pytorch_auto_revert/run_state_logger.py", line 159, in insert_state
    doc = self._build_state_json(repo=ctx.repo_full_name, ctx=ctx, pairs=pairs)
  File "/var/task/pytorch_auto_revert/run_state_logger.py", line 62, in _build_state_json
    if outcome.job_base_name:
```

Optional field was accessed unconditionally on the outcome. Removed
accessor, as field is not used.

Added more explicit logging & tests.

v20251006-140834

[AUTOREVERT] Improvements on CLI interface (#7297)

In order to avoid any confusion, and adding a few improvements after
discussion, I made significant changes to the CLI parsing for
autorevert:

1 - Adding the proper sequence parsing for CLI parameters: env variables
< explicit cli options;
2 - To clear out any possible misunderstanding on how to use the CLI,
--dry-run option is now *explicitly incompatible* with providing actions
specifications via either environment variables or cli options;
3 - As much as possible (in between reason) the code for parsing
environment variables and making decisions have been centralised and
reused;

---------

Signed-off-by: Jean Schmidt <[email protected]>