mdebray/tvl-depot
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

README: document current runner logic in detail

Browse source
This commit is contained in:
Florian Klink 2019-12-04 11:42:06 +01:00
parent 2a4d5516e4
commit 0ec77274f5
1 changed files with 51 additions and 20 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

71
README.md
View file

@ -6,41 +6,72 @@ instance, ensuring they still pass CI.
In a usual gerrit setup with a linear master history, different developers In a usual gerrit setup with a linear master history, different developers
await CI feedback on a rebased changeset, then one clicks submit, and await CI feedback on a rebased changeset, then one clicks submit, and
effectively makes everybody else rebase again. `gerrit-queue` is meant to effectively makes everybody else rebase again. `gerrit-queue` is meant to
remove these races to master. Developers can add a specific tag to a changeset, remove these races to master.
and if all preconditions are met (passing CI and passing Code Review),
`gerrit-queue` takes care of rebasing and submitting it to master. Developers can add a specific tag `submit_me` to all changesets in a series,
and if all preconditions on are met ("submittable" in gerrit speech, this
usually means passing CI and passing Code Review), `gerrit-queue` takes care of
rebasing and submitting it to master
## How it works ## How it works
Gerrit only knows about Changesets (and some relations to other changesets), Gerrit only knows about Changesets (and some relations to other changesets),
but usually developers think in terms of multiple changesets. but usually developers think in terms of multiple changesets.
### The Fetching Phase ### Fetching changesets
`gerrit-queue` fetches all changesets from gerrit, and tries to identify these `gerrit-queue` fetches all changesets from gerrit, and tries to identify these
chains of changesets. We call them `Series`. All changesets need to have strict chains of changesets. We call them `Series`. All changesets need to have strict
parent/child relationships to be detected. parent/child relationships to be detected (so if only half of the stack gets
rebased by the Gerrit Web interface, these are considered individual series.
Series are ordered by the number of changesets in them. This ensures longer Series are sorted by the number of changesets in them. This ensures longer
series are merged faster, and less rebases are triggered. In the future, this series are merged faster, and less rebases are triggered. In the future, this
might be extended to other metrics. might be extended to other metrics.
### The Submit Phase ### Submitting changesets
We loop over all series. If all changesets of a given series pass the required The submitqueue has a Trigger() function, which gets periodically executed.
preconditions (passing CI, passing Code Review, autosubmit Tag) and it's
rebased on top of the current destination branch's HEAD, it can be submitted.
The current serie is removed from the list of series, and we update our HEAD to It can keep a reference to one single serie across multiple runs. This is
the commit ID of the last commit in the submitted series. necessary if it previously rebased one serie to current HEAD and needs to wait
some time until CI feedback is there. If it wouldn't keep that state, it would
pick another series (with +1 from CI) and trigger a rebase on that one, so
depending on CI run times and trigger intervals, if not keepig this information
it'd end up rebasing all unrebased changesets on the same HEAD, and then just
pick one, instead of waiting for the one to finish.
### The Rebase Phase The Trigger() function first instructs the gerrit client to fetch changesets
We loop over all remaining series. The first one matching the required and assemble series.
preconditions is rebased on top of the (advanced HEAD) (and if this fails, we If there is a `wipSerie` from a previous run, we check if it can still be found
skip and keep trying with the next Series one after another). in the newly assembled list of series (it still needs to contain the same
number of series. Commit IDs may differ, because the code doesn't reassemble a
`wipSerie` after scheduling a rebase.
If the `wipSerie` could be refreshed, we update the pointer with the newly
assembled series. If we couldn't find it, we drop it.
Now, we enter the main for loop. The first half of the loop checks various
conditions of the current `wipSerie`, and if successful, does the submit
("Submit phase"), the second half will pick a suitable new `wipSerie`, and
potentially do a rebase ("Pick phase").
These three phases are designed to be stateless, and currently triggered once #### Submit phase
per run. We check if there is an existing `wipSerie`. If there isn't, we immediately go to
This is supposed to be moved to a more reactive model, so that the submit phase the "pick" phase.
is triggered by Webhooks for CI feedback, and we don't rebase too often
The `wipSerie` still needs to be rebased on `HEAD` (otherwise, the submit queue
advanced outside of gerrit), and should not fail CI (logical merge conflict) -
otherwise we discard it, and continue with the picking phase.
If the `wipSerie` still contains a changeset awaiting CI feedback, we `return`
from the `Trigger()` function (and go back to sleep).
If the changeset is "submittable" in gerrit speech, and has the necessary
submit queue tag set, we submit it.
#### Pick phase
The pick phase finds a new `wipSerie`. It'll first try to find one that already
is rebased on the current `HEAD` (so the loop can just continue, and the next
submit phase simply submit), and otherwise fall back to a not-yet-rebased
serie. Because the rebase mandates waiting for CI, the code `return`s the
`Trigger()` function, so it'll be called again after waiting some time.
## Compile and Run ## Compile and Run
```sh ```sh