Skip to content

Run AIR preflight validation in parallel with artifact staging - #7368

Merged
stevenchen-db merged 8 commits into
mainfrom
air/parallel-preflight
Nov 7, 2026
Merged

stevenchen-db merged 8 commits into
mainfrom
air/parallel-preflight

Conversation

@stevenchen-db

@stevenchen-db stevenchen-db commented Nov 5, 2026 •

Copy link
Copy Markdown
Contributor

Changes

databricks air run waited for the server-side config:validate call before staging launch artifacts. Validation doesn't read the uploaded files, so it now runs concurrently with mkdirs and the uploads (errgroup), and submit waits for both.

  • A rejected config is never submitted. The first failure cancels the other side and is the error reported; plain-tar snapshot packaging now stops on cancellation too.
  • Usage-policy lookup and the experiment directory now run after validation, so field errors keep precedence and a rejected config never creates the experiment directory.
  • Trade-off: a rejected config can leave an empty launch directory under .air/cli_launch/, and a bad usage_policy_name is reported after the upload.

Tests

  • New tests: validation overlaps staging; a validation failure cancels staging promptly; field errors beat usage-policy errors; cancelled plain-tar packaging stops.
  • go test ./cmd/air (touched tests with -race), TestAccept/air, lint pass.
  • e2-dogfood, minimal A10 run, 4 interleaved runs per build: uploads finish ~75ms sooner (median ~320ms vs ~395ms); total ~853ms vs ~944ms. Rejected configs (field error, field + bad policy, bad policy) behave as above with no run created.

This pull request and its description were written by Isaac.

`databricks air run` used to wait for the server-side config:validate
call before staging the launch artifacts. The backend check doesn't read
the uploaded files, so run it concurrently with usage-policy resolution,
the experiment directory check, mkdirs, and the uploads. Submission still
waits for both, and a failure on either side cancels the other, so an
invalid config is never submitted (it may now leave staged artifacts).

The "validation unavailable" warning is printed after staging so it
doesn't interleave with the staging spinner.

Co-authored-by: Isaac <no-reply@databricks.com>
@stevenchen-db
stevenchen-db requested review from a team as code owners November 5, 2026 00:09
@stevenchen-db
stevenchen-db requested a review from rclarey November 5, 2026 00:09
Co-authored-by: Isaac <no-reply@databricks.com>
@github-actions github-actions Bot added the AIR Databricks AI Runtime CLI label Nov 5, 2026
@eng-dev-ecosystem-bot

eng-dev-ecosystem-bot commented Nov 5, 2026 •

Copy link
Copy Markdown
Collaborator

✅ Integration test report

Commit: c6b050c

Run: 42264716121

Env ✅​pass 🙈​skip Time
✅​ aws linux-2core-8gb 271 15 4:09
✅​ aws-windows-latest-4core-16gb 273 13 3:20
✅​ azure linux-2core-8gb 270 15 3:54
✅​ azure-windows-latest-4core-16gb 272 13 3:06
✅​ gcp linux-2core-8gb 271 15 3:55
✅​ gcp-windows-latest-4core-16gb 273 13 3:18
Top 3 slowest tests (at least 2 minutes):
duration env testname
3:13 aws-windows-latest-4core-16gb TestAccept
3:11 gcp-windows-latest-4core-16gb TestAccept
3:00 azure-windows-latest-4core-16gb TestAccept

stevenchen-db and others added 6 commits November 5, 2026 04:31
Start validation in a goroutine before staging and join it before submit,
instead of extracting staging into a helper, so the diff doesn't move
existing code.

Co-authored-by: Isaac <no-reply@databricks.com>
Validation and staging now share a cancellable context: a validation
failure cancels in-flight uploads and is reported instead of the
cancellation, and a staging failure cancels in-flight validation.

Co-authored-by: Isaac <no-reply@databricks.com>
Start validation right before the upload and run the usage policy lookup
and experiment directory creation after it, so validation field errors
keep precedence and a rejected config never creates the experiment
directory, as before.

Co-authored-by: Isaac <no-reply@databricks.com>
errgroup already returns the first error and cancels the other side, so
the manual goroutine, channel, and deferred error override are not needed.

Co-authored-by: Isaac <no-reply@databricks.com>
# Conflicts:
#	cmd/air/validateconfig_test.go
A failed config validation now cancels artifact staging, but plain-tar
packaging ignored its context and archived the whole tree before the
validation error surfaced. Fail tar writes once the context is done.

Also bound the cancellation test so a regression fails instead of
hanging, and update the submitWorkload doc comment for the new order.

Co-authored-by: Isaac <no-reply@databricks.com>
@stevenchen-db
stevenchen-db added this pull request to the merge queue Nov 7, 2026
Merged via the queue into main with commit 21db1be Nov 7, 2026
28 checks passed
@stevenchen-db
stevenchen-db deleted the air/parallel-preflight branch November 7, 2026 05:19
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

AIR Databricks AI Runtime CLI

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants