Summary:
Fixes T4810. When a buildable completes, make an effort to update the corresponding object with a success or failure message. Commits don't support this yet, but revisions do.
{F144614}
Test Plan:
- Used `bin/harbormaster build` and `bin/harbormaster update` to run a pile of builds.
- Tried good/bad builds.
- Sent some normal mail to make sure the mail reentrancy change didn't break stuff.
Reviewers: btrahan
Reviewed By: btrahan
Subscribers: epriestley
Maniphest Tasks: T4810
Differential Revision: https://secure.phabricator.com/D8803
Summary:
Ref T4809. Currently, buildables have a status field but nothing populates it. Populate it:
- When builds change state, update the Buildable state.
- Use the new Buildable state on the web UI.
- Return the new Buildable state from Conduit.
To make it easier to debug/test this:
- Provide `bin/harbormaster update Bxxx ...` to force foreground update of a Buildable.
Test Plan:
- Used `bin/harbormaster update Bxxx --force --trace` to update buildables.
- Looked at buidlable list, saw statuses reported properly.
- Used Conduit to read statuses.
Reviewers: btrahan
Reviewed By: btrahan
Subscribers: epriestley
Maniphest Tasks: T4809
Differential Revision: https://secure.phabricator.com/D8799
Summary:
This hooks up all the pieces of the build pipeline so `harbormaster.sendmessage` actually works. Particularly:
- Candidate build steps (i.e., those which interact with external systems) can now "Wait for Message". This pauses them indefinitely when they complete, until something calls `harbormaster.sendmessage`.
- After processing a target, we check if we should move it to PASSED or WAITING.
- Before updating a build, we move WAITING targets with pending messages to either PASSED or FAILED.
- I added an explicit "Building" state, which doesn't affect workflows but communicates more information to human users.
A big part of this is avoiding races. I believe we get the correct behavior no matter which order events occur in:
- We update builds after targets complete and after we receive messages, so we're guaranteed to update once both these conditions are true. This means messages can't be lost (even if they arrive before a build completes).
- The minor changes to the build engine logic mean that firing additional build updates is always safe, no matter what the current state of the build is.
- The build itself is protected by a lock in the build engine.
- The target is not covered by an explicit lock, but for all states only the engine (waiting) //or// the worker (all other states) can interact with it. All of the interactions also move the target state forward to the same destination and have no other side effects.
- Messages are only consumed inside the engine lock, so they don't need an explicit lock.
Test Plan:
- Made an HTTP request wait after completion, then ran a pile of builds through it using `bin/harbormaster build` and the web UI.
- Passed and failed message-awaiting builds with `harbormaster.sendmessage`.
Reviewers: btrahan
Reviewed By: btrahan
Subscribers: epriestley, zeeg
Differential Revision: https://secure.phabricator.com/D8788
Summary: Fixes T4336. This updates the build engine to delete all artifacts when targets are being deleted. This prevents conflicts when builds are restarted.
Test Plan: Restarted a build that had a lease host step and it didn't crash.
Reviewers: epriestley, #blessed_reviewers
Reviewed By: epriestley
CC: Korvin, epriestley, aran
Maniphest Tasks: T4336
Differential Revision: https://secure.phabricator.com/D8092
Summary:
Ref T2015. Several fixes:
- `checkForCancellation()` no longer exists, and isn't relevant for resumable stops. Throw it away for now.
- Fix an issue where a build could pass even if the final step failed.
- `phlog()` exceptions so they show up in `bin/harbormaster` and the daemon logs.
- Write an exception log if a step fails.
- Add a "throw an exception" step to debug this stuff more easily.
Test Plan:
- Grepped for `checkForCancellation()`.
- Ran a failing build where the final step caused the failure.
- Observed `phlog()` in `bin/harbormaster` output.
- Observed log in web UI:
{F101168}
Reviewers: btrahan, hach-que
Reviewed By: btrahan
CC: aran
Maniphest Tasks: T2015
Differential Revision: https://secure.phabricator.com/D7935
Summary: Ref T1049. The logic in the BuildEngine is a little different from the logic on the Build itself. Make these more consistent, and make queued commands more private.
Test Plan: Restarted, stopped, and resumed a build.
Reviewers: btrahan
Reviewed By: btrahan
CC: aran
Maniphest Tasks: T1049
Differential Revision: https://secure.phabricator.com/D7897
Summary:
Ref T1049. Currently you can cancel a build, but now that we're tracking a lot more state we can stop, resume, and restart builds.
When the user issues a command against a build, I'm writing it into an auxiliary queue (`HarbormasterBuildCommand`) and then reading them out in the worker. This is mostly to avoid race messes where we try to `save()` the object in multiple places: basically, the BuildEngine is the //only// thing that writes to Build objects, and it holds a lock while it does it.
Test Plan:
- Created a plan which runs "sleep 2" a bunch of times in a row.
- Stopped, resumed, and restarted it.
Reviewers: btrahan
Reviewed By: btrahan
CC: aran, chad
Maniphest Tasks: T1049
Differential Revision: https://secure.phabricator.com/D7892
Summary:
Ref T1049. Currently, the Harbormaster worker looks like this:
foreach (step) {
run_step(step);
}
This means steps can't ever be run in parallel. Instead, split it into two workers. The "Build" worker starts things off, and basically does:
update_build();
(We could theoretically do this in the original process because it should never take very long, but since there's a lock and it's a little bit complex it seemed cleaner to separate it.)
The "Target" worker runs an individual target (like a command, or an HTTP request, or whatever), then updates the build:
run_one_step(step);
update_build();
The new `update_build()` mechanism in `HarbormasterBuildEngine` does this, roughly:
figure_out_overall_status_of_all_steps();
if (build is done) { done(); }
if (build is fail) { fail(); }
foreach (step that is ready to run) {
queue_target_worker_for_step(step);
}
So, overall:
- The part of the code that updates Builds is completely separated from the part of the code that updates Targets.
- Targets can run in parallel.
Test Plan:
- Ran a bunch of builds via `bin/harbormaster build`.
- Ran a bunch of builds via web UI.
Reviewers: btrahan
Reviewed By: btrahan
CC: aran
Maniphest Tasks: T1049
Differential Revision: https://secure.phabricator.com/D7890