This PR attempts to add some automation around assigning release
management. The PR adds a file `release/rotation`; each week, the
updated CI cron job will:
- Open a PR for the new release [as current].
- Assign the first user in the file to that PR.
- Add the Standard-Change label to the PR.
- Start the build for that PR [as current].
- Open a new PR that rotates the `release/rotate` file, i.e. pushes back
the first line to the end of the file.
This PR also adds mentions of the "release handler" (the first line of
`release/rotation`) to the various messages we send to Slack along the
release process.
The initial state of the `release/rotation` file has been created by
listing all the volunteers (Language team, Application Runtime team, as
well as @SamirTalwar-DA and @stefanobaghino-da) and piping the file
through `shuf`. (Then I put myself at the top so I can hopefully iron
out the issues with the first attempt.)
CHANGELOG_BEGIN
CHANGELOG_END
This is running on Azure’s agents and we just call curl so there is
really no need for dev-env (as evidenced by the fact that the message
got sent despite dev-env failing).
changelog_begin
changelog_end
Based on feedback from @nickchapman-da, this PR aims at making the
release process easier by:
- Automatically opening a release PR on Wednesday morning. The goal here
is that by the time we start working, there is a release already
built, so we save about an hour on waiting for that. This obviously
doesn't help with ad-hoc releases.
- On a release PR build, posting to Slack when the release is ready to
merge.
- On a release master build, posting to Slack when a release is ready to
be tested.
My hope is that this makes the release process less tedious. This is not
trying to address the actual release testing, but hopefully should
reduce the annoyance of having to constantly go and check if the release
is ready.
CHANGELOG_BEGIN
CHANGELOG_END
While closely following the 1.3 release through our pipeline to check
that #6709 worked as expected, I realized that the automatically-created
PR does not start the normal tests either, presumably because it's been
opened by a bot. The bot doe shave write access to the repo (obviously,
as it can create the PR in the first place), but somehow that doesn't
seem to count as a PR with write access for Azure.
So this PR adds the normal test run too, so we don't need to manually
say `/azp run` on the PR.
CHANGELOG_BEGIN
CHANGELOG_END
This reverts commit 61e9df3eaf.
This interacts very badly with the fact that we check out old commits
for releases. While we could fix it for this particular issue, I don’t
think this buys us enough to make this worth doing and it makes it
easy to introduce issues in the future if we modify lib.sh
changelog_begin
changelog_end
This asks Azure to run a full compat test against the branch that
updates the compat test matrix, so we know it's good when we merge it
rather than the next morning.
CHANGELOG_BEGIN
CHANGELOG_END
This PR fixes a few things with the script that automatically updates
the compat test matrix on release, based on its use over the past two
weeks. Specifically:
- Send an error message to Slack in case this job fails. Previously,
failures here were silent.
- Add an exponential backoff strategy to wait for the artifact to be
available on Maven. Previously, the update script just failed.
- Allow for rerunning on the same machine after failure by removing the
branch if it exists.
- Fix the commit message to include proper newlines instead of literal
`\n`'s.
CHANGELOG_BEGIN
CHANGELOG_END
This bumps the timeout of the compat tests on PRs to 360 minutes
matching other jobs on a PR (we mainly hit this if ghc-lib is rebuilt)
and the timeout on the daily jobs to 720 minutes (we hit this if
_everything_ is rebuilt).
I am slightly worried about the timeout on the daily job. After having
taken a look at it, there are a few reasons how we ended up here:
1. We started including more tests, e.g., sandbox-classic. Not much we
can do here, those tests are useful.
2. We have a very large number of snapshots for 1.3.0. There are a few
reasons for this:
1. Timing: We branched off early for the 1.2.0 release so the first
snapshot for 1.3 was on June 3th. For 1.4 it looks like the first
snapshot will be on July 15th so that’s roughly 2 extra
snapshots just due to timing.
2. Additional snapshots: We had one broken snapshot due to a broken
VSCode extension that we didn’t delete (probably not worth doing
at this point). We also had to backport to an old snapshot which
resulted in another extra snapshot. We also had one extra
snapshot which was supposed to be the RC but wasn’t since the
ANF revert needed to go in.
The only thing that is clearly useless is the one broken snapshot
but that doesn’t change things that much. I see 2 orthogonal
options for improving this assuming we agree that the current
runtime is worryingly high.
1. Prune snapshots more aggressively, e.g., only include the last 3
snapshots. That’s a pretty arbitrary decision but it would
enforce a hard limit.
2. Reduce test combinations. E.g., only test snapshots vs stable
releases but not snapshots vs snapshots.
3. We end up forcing a full build quite frequently. Here are just 2
examples of how we’ve done that so far.
1. Upgrade rules_haskell. Basically all tests are run by a Haskell
binary so this forces a full rebuild.
2. Change runfiles of `daml`.
I don’t think there is much we can do about 1 or 3 which leaves us
with 2. One not entirely unreasonable option is to just do nothing. We
did have periods where things went pretty smoothly for the most part
and each month we reset to a much smaller number of releases (we also
have to start throwing out old stable releases at some
point). Otherwise reducing the number of test combinations seems the
most promising option to me.
changelog_begin
changelog_end
We take our own libraries from latest_stable_version which changed but
we did not rerun pinning which meant that this did not get updated.
changelog_begin
changelog_end
For a while now we've had errors along the line of
```
FATAL: could not create shared memory segment: No space left on device
DETAIL: Failed system call was shmget(key=5432001, size=56, 03600).
HINT: This error does *not* mean that you have run out of disk space.
It occurs either if all available shared memory IDs have been taken, in
which case you need to raise the SHMMNI parameter in your kernel, or
because the system's overall limit for shared memory has been reached.
The PostgreSQL documentation contains more information about
shared memory configuration.
child process exited with exit code 1
```
on macOS CI nodes, which we were not able to reproduce locally. Today I
managed to, sort of by accident, and that allowed me to dig a bit
further.
The root cause seems to be that PostgreSQL, as run by Bazel, does not
always seem to properly unlink the shared memory segment it uses to
communicate with itself. On my machine, running:
```
bazel test -t- --runs_per_test=100 //ledger/sandbox:conformance-test-wall-clock-postgresql
```
and eyealling the results of
```
watch ipcs -mcopt
```
I would say about one in three runs leaks its memory segment. After much
googling and some head scratching trying to figure out the C APIs for
managing shared memory segments on macOS, I kind of stumbled on a
reference to `pcirm` in a comment to some low-ranking StackOverflow
answer. It looks like it's working very well on my machine, even if I
run it while a test (and therefore an instance of pg) is running. I
believe this is because the command does not actually remove the shared
memory segments, but simply marks them for removal once the last process
stops using it. (At least that's what the manpage describes.)
CHANGELOG_BEGIN
CHANGELOG_END
This PR adds an extra post-release job to CI that will run the
[`compatiblity/update-versions.sh`][0] script and open a PR with the
result.
[0]: cb82a8d6be/compatibility/update-versions.sh
CHANGELOG_BEGIN
CHANGELOG_END
fixes#6384
For now this keeps the GCP bucket as well. I would suggest to keep
that for 1.3 and drop it in 1.4 but I don’t feel particularly strongly
about this so I’m also happy to drop it now.
changelog_begin
- [SDK] The JSON API and DAML on SQL (sandbox-classic) are now
published as fat JARs to github releases. The GCP bucket that
contained the fat JARs will not receive releases > 1.3.
changelog_end
Nix now requires -L, I’ve gone ahead and just normalized everything to
use -sfL which we were already using in one place.
changelog_begin
changelog_end
automated ghc-lib build
This PR aims at automating the build of ghc-lib. The current process
still has a few manual steps; it needs to be updated because Bintray is
going away, so this seemed like a good opportunity to fully automate it.
This works like the "patch bazel on Windows" jobs: the filename will
contain a hash of the `ci/da-ghc-lib` folder, and the job will run only
if the corresponding filename does not yet exist on the GCS bucket. PRs
aiming at changing the ghc-lib version will need to run twice: once to
create the artifacts, and once to change the `stack-snapshot.yaml` file
to match.
CHANGELOG_BEGIN
CHANGELOG_END
trigger all releases from master
The 1.1.0 release went wrong and we had to trash it and release 1.1.1
instead. This is an attempt at identifying and correcting the root
cause behind that incident.
To understand the situation, we need to know how releases worked before
1.0. We had a one-line file called `LATEST` that specifies the git SHA and
version tag for the latest release. A change to that file triggered a
release with the specified release tag, built from the source tree of
the specified commit. The `LATEST` file looked something like:
```
f050da78c9 1.0.0-snapshot.20200411.3905.0.f050da78
```
To mark a release as stable, we would change it to look like this:
```
f050da78c9 1.0.0
```
i.e. simply drop the `-snapshot...` suffix. Even though the commit (and
thus the entire source tree we build from) is the same, we would need to
rebuild almost all of our release artifacts, as they embed the version
tag in various places and ways. That worked well as long as we could
assume we were doing trunk-based development, i.e. all releases would
always come from the same (`master`) branch.
When we released 1.0, and started work on 1.1, we had a few bug reports
for 1.0 that we decided should be resolved in a point release. We
decided that the best way to handle that would be to have a branch
starting on the release commit for 1.0, and then backport patches from
`master` to that branch. We adapted our build process to also watch the
`release/1.0.x` branch and, in particular, trigger a new release build if
the `LATEST` file in that branch changed. That worked well.
The plan going forward was to keep doing regular snapshot releases from
the `master` branch, and create support, point releases ("patch" releases
in semver) from dedicated branches.
On April 30, we made a snapshot release as an RC for 1.1.0, by changing
the `LATEST` file in the `master` branch. That release was built on commit
681c862d. On May 6, we decided to take a new snapshot as the RC for
1.1.0; we changed `LATEST` in `master` to designate 7e448d81 as the new
latest release.
On May 11, we noticed an issue that broke our builds. Without going into
details, an external artifact we depend on had changed in incompatible
ways. After fixing that on `master`, we reasoned that this would also
break the build of the final 1.1.0 release if we just tried to build
7e448d81 again. But as the target release date was May 13, we did not
want to take a new snapshot after that fix, as that would have included
one more week of work in the release, and given us no time to test it.
So we did what we did for the 1.0 branch, as it had worked well: we
created a branch that forked from `master` at commit 7e448d81 and called
it `release/1.1.x`, then cherry-picked the one fix to our build process to
work around the broken download. When the time came to make the final
1.1.0 build on May 13, we naturally picked the `LATEST` file from the
`release/1.1.x` branch and dropped the `-snapshot...` suffix. Importantly,
we did not need to update the target commit to include the "broken
download" fix as, in the meantime, the internet had fixed itself, and we
thus reasoned we should go for the exact code of the RC rather than
include an unnecessary, albeit seemingly harmless, change.
Everything went well with the release process. Tests went well too. Then
we got a report that an application that worked against the latest RC
broke with the final 1.1.0. The issue was that we had built the wrong
commit: by branching off at the point of the _target_ commit for the
latest snapshot, we did not have the change to the `LATEST` file that
designated that commit as the target. So the `LATEST` file in
`release/1.1.x` was still pointing to 681c862d.
I believe the root cause for this issue is the fact that we have
scattered our release process over multiple branches, meaning there is
no linear history of what was released and we are relying on people
being able to mentally manage multiple timelines. Therefore, I propose
to fix our release process so this should not happen again by
linearizing the release process, i.e. getting back to a situation where
all releases are made from a single branch, `master`.
Because we do want to be able to release _for_ multiple release branches
(to provide backports and bugfixes), we still need some way to
accommodate that. Having a single `LATEST` file in the same format as
before would not really work well: keeping track of interleaved release
streams on a single file would not really be easier than keeping track
of multiple branches.
My proposed solution is to instead have a multiline LATEST file, so that
all the release branch "tips" can be observed at the same time, and, as
long as we take care to only advance one release branch at a time, we
can easily keep track of each of them. This is what this PR does.
This required a few changes to our release process. Most notably:
- Obviously, as this is the main point of this PR, the build process has
once again been restricted to only trigger new releases from the
`master` branch.
- As our CI machinery cannot easily be made to produce multiple releases
from a single build, the `check_for_release` step will only recognize
a commit as a release trigger if it changes a single line in the
`LATEST` file. This restriction comes in addition to the existing one
that a release commit is only allowed to change either just the
`LATEST` file or both the `LATEST` and
`docs/source/support/release-notes.rst` files.
- The docs publication process has been changed to update _all_
published versions to display the _latest_ release notes page. This
means that the release notes page will always show you all published
versions, regardless of which version of the documentation you're
looking at. This also means that interleaving release notes correctly on
that page is a manual exercise.
- As per the intention of the new process, the `LATEST` file has been
updated to contained all existing post-1.0 stable releases. It should
also include all existing snapshot releases should we have more than one
at a time (say, should we discover an issue with 1.1.1 that required us
to work on a 1.1.2).
- The `release.sh` script has been dramatically simplified as I felt it
was trying to do too much and porting its existing functionality to a
multi-line `LATEST` file would be too hard.
CHANGELOG_BEGIN
CHANGELOG_END
Note: this is beta-level software. See documentation for the precise
guarantees this does and does not come with. (Documentation does not
exist at the time of opening this PR, but should exist by the time the
first version of this gets published.)
CHANGELOG_BEGIN
- We now publish Sandbox Next as an **ALPHA** standalone jar.
- We now publish the HTTP JSON API as a standalone jar.
CHANGELOG_END
This is the first part of #5700
It adds tests that build create-daml-app using `daml build` and then
run the codegen and build the UI. Contrary to our main tests these
also run on Windows. This is actually reasonably simple by first
building the typescript libraries on Linux and then downloading them
on Windows.
There are two parts that are still missing from the tests in the main
workspace:
1. Building the extra feature. This should be fairly easy to add.
2. Running the pupeeter tests. At least MacOS and Linux should be
reasonably easy. I don’t know what horrors Windows will throw at
us. This step is what actually makes this a compatibility
test. Currently it doesn’t actually launch Sandbox and the JSON API.
Since this PR is already pretty large, I’d like to tackle those things
separately.
changelog_begin
changelog_end
patch Bazel on Windows (ci setup)
We have a weird, intermittent bug on Windows where Bazel gets into a
broken state. To investigate, we need to patch Bazel to add more debug
output than present in the official distribution. This PR adds the basic
infrastructure we need to download the Bazel source code, apply a patch,
compile it, and make that binary available to the rest of the build.
This is for Windows only as we already have the ability to do similar
things on Linux and macOS through Nix.
This PR does not contain any intresting patch to Bazel, just the minimum
that we can check we are actually using the patched version.
CHANGELOG_BEGIN
CHANGELOG_END
This is temporary. It looks like the macOS nodes are dead; @nycnewman is
looking into it, but in case he doesn't fix it in time, at least we
have a backup plan so we're not completely blocked on Monday.
CHANGELOG_BEGIN
CHANGELOG_END
add default machine capability
We semi-regularly need to do work that has the potential to disrupt a
machine's local cache, rendering it broken for other streams of work.
This can include upgrading nix, upgrading Bazel, debugging caching
issues, or anything related to Windows.
Right now we do not have any good solution for these situations. We can
either not do those streams of work, or we can proceed with them and
just accept that all other builds may get affected depending on which
machine they get assigned to. Debugging broken nodes is particularly
tricky as we do not have any way to force a build to run on a given
node.
This PR aims at providing a better alternative by (ab)using an Azure
Pipelines feature called
[capabilities](https://docs.microsoft.com/en-us/azure/devops/pipelines/agents/agents?view=azure-devops&tabs=browser#capabilities).
The idea behind capabilities is that you assign a set of tags to a
machine, and then a job can express its
[demands](https://docs.microsoft.com/en-us/azure/devops/pipelines/process/demands?view=azure-devops&tabs=yaml),
i.e. specify a set of tags machines need to have in order to run it.
Support for this is fairly badly documented. We can gather from the
documentation that a job can specify two things about a capability
(through its `demands`): that a given tag exists, and that a given tag
has an exact specified value. In particular, a job cannot specify that a
capability should _not_ be present, meaning we cannot rely on, say,
adding a "broken" tag to broken machines.
Documentation on how to set capabilities for an agent is basically
nonexistent, but [looking at the
code](https://github.com/microsoft/azure-pipelines-agent/blob/master/src/Microsoft.VisualStudio.Services.Agent/Capabilities/UserCapabilitiesProvider.cs)
indicates that they can be set by using a simple `key=value`-formatted
text file, provided we can find the right place to put this file.
This PR adds this file to our Linux, macOS and Windows node init scripts
to define an `assignment` capability and adds a demand for a `default`
value on each job. From then on, when we hit a case where we want a PR
to run on a specific node, and to prevent other PRs from running on that
node, we can manually override the capability from the Azure UI and
update the demand in the relevant YAML file in the PR.
CHANGELOG_BEGIN
CHANGELOG_END
This PR changes the notify_user job to not run when the job has been
canceled, which happens mostly when we push new code.
Not sure how I failed to see the `canceled` function in the past, but
this does seem to do exactly what we want.
CHANGELOG_BEGIN
CHANGELOG_END
This PR separates the "last known valid perf test" commit from the
"baseline speedy implementation" commit. It is important for the perf
test to be meaningful that the changes between those two commits are
benign, say minor API adjustments, so that the perf measurement remains
meaningful.
This also adds a check on merging to master that tells Slack if the perf
test has changed and the `test_sha` file needs updating. The Slack
message is conditional on the current commit to avoid excessive noise.
CHANGELOG_BEGIN
CHANGELOG_END
Currently, there are quite a few releases that are lacking the
Standard-Change label, even though they did publish artifacts. This
makes our SOC2-compliance tracking a bit harder. For the past two
months, I have manually added the label after-the-fact while preparing
the monthly compliance report, but that doesn't seem like a great
solution.
This PR changes the release process to be more optimistic: assume the
release is going to succeed by putting in the label immediately, and
then (optionally) removing it if the release fails.
Note that the label should only be removed in the rare case where the
release was merged into master but somehow did not produce any artifact.
This can only happen if the Linux build fails quite early, which as far
as I know only happened once over the past two months when we had the
release notes race condition.
CHANGELOG_BEGIN
CHANGELOG_END
At the moment, collect_build_data will wait for the Windows
compatibility test to have "finished", but doesn't check its return
status. This means two things:
1. Should the compatibility test end without a success or error (e.g.
communication broken between Azure and the node), the option to rerun
failed jobs will not appear, as there will be no failed job.
2. The subsequent notify_user step will ignore failures in the
compatibility_windows job when reporting to Slack, making for
confusing reports.
CHANGELOG_BEGIN
CHANGELOG_END
* Make compat tests work on windows
This required some changes to the daml_sdk rule since the read-only
installation by the assistant breaks Bazel completely. We could only
apply those changes on Windows but I think I prefer the consistency
across platforms here over trying to stay close to how the SDK is
installed on user machines given that the SDK installation is not
something we’ve had issues with.
I’ve excluded the postgresql tests for now. I don’t expect them to be
particularly hard to fix but I’ve already spent almost 2 days on this
and having some tests run on Windows seems like a clear improvement
over running no tests on Windows :)
changelog_begin
changelog_end
* Remove todo
changelog_begin
changelog_end
This is a minor, cosmetic change. Note that all our references to
releases are based on tags, and do not depend on the release title. This
is evidenced by the fairly random titles we used to have before the
title was set by CI, see e.g.
[0.13.35](https://github.com/digital-asset/daml/releases/tag/v0.13.53).
CHANGELOG_BEGIN
CHANGELOG_END
This PR extends the existing Linux compatibility tests to run on macOS
too. Fixes#5692.
CHANGELOG_BEGIN
CHANGELOG_END
Co-authored-by: Moritz Kiefer <moritz.kiefer@purelyfunctional.org>
* Extend compatibility tests to DAML on SQL
This feels a bit hacky since the runfiles don’t work quite like I
would expect them to but it’s at least not more hacky than what we do
for the head-based tests we currently have.
Progress towards #5695
changelog_begin
changelog_end
* Fix runfiles with more bash
changelog_begin
changelog_end
* remove redundant port options
changelog_begin
changelog_end
* Create fewer sandbox targets
changelog_begin
changelog_end
* Apply suggestions from code review
Co-Authored-By: Andreas Herrmann <42969706+aherrmann-da@users.noreply.github.com>
* Fix runfiles snippet
changelog_begin
changelog_end
Co-authored-by: Andreas Herrmann <42969706+aherrmann-da@users.noreply.github.com>
This is a first step towards testing cross-version
compatibility. It doesn’t actuall do much yet but hopefully it should
be easier to parallelize once we have the initial boilerplate in place
so ideally I’d like to address most missing things and issues in
separate PRs.
changelog_begin
changelog_end
As currently written, the check will compare the current commit with the
base commit the branch was forked from. The intention was to only list
the changes from the current PR, and to make the check work for PRs
against release branches (the previous version of the heck always
compared to master, regardless of what branch the PR was targeting).
However, this does not work as expected because the "current commit" is
not the tip of the PR, but the merge commit supplied by GitHub.
Therefore, the diff here will include not only the changes in this PR,
but also all the changes that happened on the target branch since
forking. This is not an issue if the PR is properly rebased, but that's
hard to control in a world where other people work too.
This PR corrects this by explicitly computing the diff between the fork
point on the target branch and the tip of the PR, ignoring the
currently-checked-out commit.
CHANGELOG_BEGIN
CHANGELOG_END
This commit aims at enabling future patch releases; it is the
master-branch equivalent of #5569 (applied to the 1.0 release branch).
The only change between the two changelogs should be that this one also
changes the docs cron so it can find the trigger commits for patch
releases.
CHANGELOG_BEGIN
CHANGELOG_END
I recently noticed that the `check_for_release` and
`check_standard_change_label` jobs do not currently report their
runtime, so including them in the build data is a bit moot (we always
get `""` for all three values). Given that they usually run in under 3
seconds, I've decided the best way to fix this is to remove them from
the build data, rather than add the required steps to collect their
build times.
CHANGELOG_BEGIN
CHANGELOG_END
Looking at the behaviour of `succeededOrFailed`, it looks like it does
not do what we want at all: both steps now only run on failures. My
current hypothesis is that `write_ledger_dump` being skipped switches
the state of the last job to something that is neither success nor
failure.
It would be really nice if Azure had a way to detect cancellation. :(
CHANGELOG_BEGIN
CHANGELOG_END
Over the past three weeks or so, I have not seen a single case where the
"get commit" step failed erroneously, i.e. all failures were genuine
pushes, which we don't care about. Therefore, I hacve decided to remove
the `tell_gary` code, as it is long, a bit hairy, and duplicated.
In the meantime I've also discovered there actually is a way to tell
Azure not to run these steps on a canceled build, which I believe makes
sense, so I've added that.
CHANGELOG_BEGIN
CHANGELOG_END
Currently, on a release commit on master, if the commit fails, we get
the message from the target PR, which is confusing. This should
(hopefully; it's a bit hard to test as it would require setting up a
release PR that succeeds but fails on master) get us the title of the
release commit, which hopefully will be less confusing.
CHANGELOG_BEGIN
CHANGELOG_END
Currently, on Linux, after the normal build, we try running the release
script (in "dry run" mode). This is to check that the release script not
only compiles, but actually runs. To be honest I'm not entirely sure why
we do that as a separate step (i.e. why does `bazel test //...` not give
us confidence about this script?), but the point of this PR is that,
while there may be some benefit in running this script on normal PRs to
check that we have not broken the release step, there is absolutely no
point in running it _on a release build_, i.e. right after we've used
the same script in "real" ("wet run"? 🤔) mode.
CHANGELOG_BEGIN
CHANGELOG_END
This has been running for a few days now and while I have seen a bunch
of these cases, I have not once received a message with a BACKOFF value
different from 512. This means that, likely due to some sort of internal
caching in Azure, retrying in this case is useless and just makes the
build failure take more time, i.e. more time before we can rerun.
Rerunning does usually solve it, though.
I have also noticed that we still get these notifications when the job
has been canceled, which usually means the user has force-pushed (in
which case it makes sense that the commit is no longer available). I'm
not sure we can detect this, but I take this opportunity to print the
JobStatus just in case.
CHANGELOG_BEGIN
CHANGELOG_END
Somehow, in the current setup, the publish steps do not get executed on
master. This is what Azure reports:
```
Evaluating: and(succeeded(), eq('$(is_release)', 'true'),
eq(variables['Build.SourceBranchName'], 'master'), eq('linux', 'linux'))
Expanded: and(True, eq('$(is_release)', 'true'),
eq(variables['Build.SourceBranchName'], 'master'), eq('linux', 'linux'))
Result: False
```
So it looks like, in the condition, `${{parameters.is_release}}`
evaluates to the literal string `$(is_release)`. If we look at the point
of invocation of the ~function~ template, we can see:
```
- template: ci/build-unix.yml
parameters:
release_tag: $(release_tag)
name: 'linux'
is_release: $(is_release)
```
so it does not seem completely crazy. However, according to the
documentation, we should expect that to be replaced by the value of the
corresponding variable, as per:
```
variables:
release_sha: $[ dependencies.check_for_release.outputs['out.release_sha'] ]
release_tag: $[ coalesce(dependencies.check_for_release.outputs['out.release_tag'], '0.0.0') ]
trigger_sha: $[ dependencies.check_for_release.outputs['out.trigger_sha'] ]
is_release: $[ dependencies.check_for_release.outputs['out.is_release'] ]
```
What's interesting here is that, within `build-unix.yml`, we are also
using `release_tag` in the exact same way:
```
- bash: ./build.sh "_$(uname)"
displayName: 'Build'
env:
DAML_SDK_RELEASE_VERSION: ${{parameters.release_tag}}
```
and this time output from the build seems to show the value being
correctly substituted:
```
damlc - Compiler and IDE backend for the Digital Asset Modelling
Language
SDK Version: 0.13.55-snapshot.20200226.3266.d58bb459
Usage: <interactive> COMMAND
Invoke the DAML compiler. Use -h for help.
```
My current guess is that the (undocumented, as far as I can tell)
evaluation order is as follows:
1. In the template, syntactically replace all the parameters.
2. In the job definition, replace the call to the template with the code
of the template. So it is as if we had written the template directly in
the `azure-pipelines.yml` file, with `$(release_tag)` and
`$(is_release)`.
3. Run the build. When we reach the time to run this specific job,
we can evaluate the expressions for the variables and replace them in
the rest of the job.
So what is going wrong? I believe the issue is with the quotes,
preventing the substitution of `is_release`. They came directly from the
[documented
syntax](https://docs.microsoft.com/en-us/azure/devops/pipelines/process/conditions?view=azure-devops&tabs=yaml#use-a-template-parameter-as-part-of-a-condition),
but if the above evaluation order is correct, they should not be there.
There are actually two things going wrong here. The first one is that
the syntax `$()` is used to substitute a value in what Azure considers a
string. This is the case for `env` keys. However, the `condition` key
is not a string, it is an Azure "expression". Expressions have their own
evaluation rules and syntax, and in particular, `$()` is not a
substitution rule there, so when it sees `$()` in a string in an
expression (due to the quoptes), it leaves it alone.
Removing the quotes does not directly help, though, as we then end with
```
condition: eq($(is_release), 'true')
```
and `$()` is not valid syntax in an expression. The way to use variables
in an expression is `variables.name` (or `variables["name"]`, because
why have only one?).
So that means we have to pass variables to the template in different
ways depending on how they will be used. So much fun.
CHANGELOG_BEGIN
CHANGELOG_END
The existing approach is a historical accident. The reason for the
additional tarball/install jobs was that, in my original attempt, the
build steps would still build the current commit, as opposed to the
target commit.
This is not such an issue on Linux, but setting up the build environment
on Windows and macOS _again_ for no good reason is a pure waste of time
(and effort in getting it right). Now that the build steps build the
target commit (with the env var set), we can go back to the way things
were previously: just take the build products directly from the build
step.
CHANGELOG_BEGIN
CHANGELOG_END
This doesn’t seem to bring any benefit anymore on cache hits with the
reduced closure size and on cache misses it’s significantly slower.
changelog_begin
changelog_end
This removes the sample/reference implementation of kvutils
InMemoryKVParticipantState.
This used to be the only implementation of kvutils, but now with the
simplified kvutils api we have ledger-on-memory and ledger-on-sql.
InMemoryKVParticipantState was also used for the ledger dump utility,
which now uses ledger-on-memory.
* Runner now supports a multi participant configuration
This change removes the "extra participants" config and goes for consistent
participant setup with --participant.
* Run all conformance tests in the repository in verbose mode.
This means we'll print stack traces on error, which should make it
easier to figure out what's going on with flaky tests on CI.
This doesn't change the default for other users of the
ledger-api-test-tool; we just add the flag for:
- ledger-api-test-tool-on-canton
- ledger-on-memory
- ledger-on-sql
- sandbox
Fixes#4225.
CHANGELOG_BEGIN
CHANGELOG_END
The default behaviour of an Azure job that has a dependency is to only
run if the dependency has succeeded. However, that default behaviour is
overridden if there is an exmplicit `condition` attribute.
This PR restores the expected behaviour that we only try to build a
release tarball if the actual build has succeeded.
CHANGELOG_BEGIN
CHANGELOG_END
While flailing about randomly trying to reset the Windows cache
yesterday, I noticed a couple issues with the current script:
- The `fork_point` calculation is just plain broken. Somehow our `set
-euo pipefail` does not fail on subshell errors, but the existing
command is just never going to work: it looks like `git` does not
resolve refs on the `merge-base` command. It also looks like the
`--fork-point` option is not what we want. I don't know how this
happened.
- `sort` on my machine and on CI do not seem to behave the same with
respect to upper/lower case ordering. To make the script independent
of the specific sort order on the machine (probably controlled by the
locale), we now sort both the actual and the expected list.
Finally, based on the failure to recognize a release commit once merged
into master, I realized that of course computing the diff between a
commit and itself will yield an empty diff. The `git_sha` step will now
identify the "master" and "fork point" commits as the parent for a
master build.
CHANGELOG_BEGIN
CHANGELOG_END
In the current setup, we expose HEAD as the trigger commit, but that is
the merge commit with master. Since making a release takes a long time,
this merge commit is likely to not exist anymore by the time we want to
try a rerun (assuming a flaky build).
Release PRs are by definition (in the new system) independent of what's
going on on master, so we should instead take the branch commit here
when running on a PR.
CHANGELOG_BEGIN
CHANGELOG_END
Context
=======
After multiple discussions about our current release schedule and
process, we've come to the conclusion that we need to be able to make a
distinction between technical snapshots and marketing releases. In other
words, we need to be able to create a bundle for early adopters to test
without making it an officially-supported version, and without
necessarily implying everyone should go through the trouble of
upgrading. The underlying goal is to have less frequent but more stable
"official" releases.
This PR is a proposal for a new release process designed under the
following constraints:
- Reuse as much as possible of the existing infrastructure, to minimize
effort but also chances of disruptions.
- Have the ability to create "snapshot"/"nightly"/... releases that are
not meant for general public consumption, but can still be used by savvy
users without jumping through too many extra hoops (ideally just
swapping in a slightly-weirder version string).
- Have the ability to promote an existing snapshot release to "official"
release status, with as few changes as possible in-between, so we can be
confident that the official release is what we tested as a prerelease.
- Have as much of the release pipeline shared between the two types of
releases, to avoid discovering non-transient problems while trying to
promote a snapshot to an official release.
- Triggerring a release should still be done through a PR, so we can
keep the same approval process for SOC2 auditability.
The gist of this proposal is to replace the current `VERSION` file with
a `LATEST` file, which would have the following format:
```
ef5d32b7438e481de0235c5538aedab419682388 0.13.53-alpha.20200214.3025.ef5d32b7
```
This file would be maintained with a script to reduce manual labor in
producing the version string. Other than that, the process will be
largely the same, with releases triggered by changes to this `LATEST`
and the release notes files.
Version numbers
===============
Because one of the goals is to reduce the velocity of our published
version numbers, we need a different version scheme for our snapshot
releases. Fortunately, most version schemes have some support for that;
unfortunately, the SDK sits at the intersection of three different
version schemes that have made incompatible choices. Without going into
too much detail:
- Semantic versioning (which we chose as the version format for the SDK
version number) allows for "prerelease" version numbers as well as
"metadata"; an example of a complete version string would be
`1.2.3-nightly.201+server12.43`. The "main" part of the version string
always has to have 3 numbers separated by dots; the "prerelease"
(after the `-` but before the `+`) and the "metadata" (after the `+`)
parts are optional and, if present, must consist of one or more segments
separated by dots, where a segment can be either a number or an
alphanumeric string. In terms of ordering, metadata is irrelevant and
any version with a prerelease string is before the corresponding "main"
version string alone. Amongst prereleases, segments are compared in
order with purely numeric ones compared as numbers and mixed ones
compared lexicographically. So 1.2.3 is more recent than 1.2.3-1,
which is itself less recent than 1.2.3-2.
- Maven version strings are any number of segments separated by a `.`, a
`-`, or a transition between a number and a letter. Version strings
are compared element-wise, with numeric segments being compared as
numbers. Alphabetic segments are treated specially if they happen to be
one of a handful of magic words (such as "alpha", "beta" or "snapshot"
for example) which count as "qualifiers"; a version string with a
qualifier is "before" its prefix (`1.2.3` is before `1.2.3-alpha.3`,
which is the same as `1.2.3-alpha3` or `1.2.3-alpha-3`), and there is a
special ordering amongst qualifiers. Other alphabetic segments are
compared alphabetically and count as being "after" their prefix
(`1.2.3-really-final-this-time` counts as being released after `1.2.3`).
- GHC package numbers are comprised of any number of numeric segments
separated by `.`, plus an optional (though deprecated) alphanumeric
"version tag" separated by a `-`. I could not find any official
documentation on ordering for the version tag; numeric segments are
compared as numbers.
- npm uses semantic versioning so that is covered already.
After much more investigation than I'd care to admit, I have come up
with the following compromise as the least-bad solution. First,
obviously, the version string for stable/marketing versions is going to
be "standard" semver, i.e. major.minor.patch, all numbers, which works,
and sorts as expected, for all three schemes. For snapshot releases, we
shall use the following (semver) format:
```
0.13.53-alpha.20200214.3025.ef5d32b7
```
where the components are, respectively:
- `0.13.53`: the expected version string of the next "stable" release.
- `alpha`: a marker that hopefully scares people enough.
- `20200214`: the date of the release commit, which _MUST_ be on
master.
- `3025`: the number of commits in master up to the release commit
(included). Because we have a linear, append-only master branch, this
uniquely identifies the commit.
- `ef5d32b7ù : the first 8 characters of the release commit sha. This is
not strictly speaking necessary, but makes it a lot more convenient to
identify the commit.
The main downsides of this format are:
1. It is not a valid format for GHC packages. We do not publish GHC
packages from the SDK (so far we have instead opted to release our
Haskell code as separate packages entirely), so this should not be an
issue. However, our SDK version currently leaks to `ghc-pkg` as the
version string for the stdlib (and prim) packages. This PR addresses
that by tweaking the compiler to remove the offending bits, so `ghc-pkg`
would see the above version number as `0.13.53.20200214.3025`, which
should be enough to uniquely identify it. Note that, as far as I could
find out, this number would never be exposed to users.
2. It is rather long, which I think is good from a human perspective as
it makes it more scary. However, I have been told that this may be
long enough to cause issues on Windows by pushing us past the max path
size limitation of that "OS". I suggest we try it and see what
happens.
The upsides are:
- It clearly indicates it is an unstable release (`alpha`).
- It clearly indicates how old it is, by including the date.
- To humans, it is immediately obvious which version is "later" even if
they have the same date, allowing us to release same-day patches if
needed. (Note: that is, commits that were made on the same day; the
release date itself is irrelevant here.)
- It contains the git sha so the commit built for that release is
immediately obvious.
- It sorts correctly under all schemes (modulo the modification for
GHC).
Alternatives I considered:
- Pander to GHC: 0.13.53-alpha-20200214-3025-ef5d32b7. This format would
be accepted by all schemes, but will not sort as expected under semantic
versioning (though Maven will be fine). I have no idea how it will sort
under GHC.
- Not having any non-numeric component, e.g. `0.13.53.20200214.3025`.
This is not valid semantic versioning and is therefore rejected by
npm.
- Not having detailed info: just go with `0.13.53-snapshot`. This is
what is generally done in the Java world, but we then lose track of what
version is actually in use and I'm concerned about bug reports. This
would also not let us publish to the main Maven repo (at least not more
than once), as artifacts there are supposed to be immutable.
- No having a qualifier: `0.13.53-3025` would be acceptable to all three
version formats. However, it would not clearly indicate to humans that
it is not meant as a stable version, and would sort differently under
semantic versioning (which counts it as a prerelease, i.e. before
`0.13.53`) than under maven (which counts it as a patch, so after
`0.13.53`).
- Just counting releases: `0.13.53-alpha.1`, where we just count the
number of prereleases in-between `0.13.52` and the next. This is
currently the fallback plan if Windows path length causes issues. It
would be less convenient to map releases to commits, but it could still
be done via querying the history of the `LATEST` file.
Release notes
=============
> Note: We have decided not to have release notes for snapshot releases.
Release notes are a bit tricky. Because we want the ability to make
snapshot releases, then later on promote them to stable releases, it
follows that we want to build commits from the past. However, if we
decide post-hoc that a commit is actually a good candidate for a
release, there is no way that commit can have the appropriate release
notes: it cannot know what version number it's getting, and, moreover,
we now track changes in commit messages. And I do not think anyone wants
to go back to the release notes file being a merge bottleneck.
But release notes need to be published to the releases blog upon
releasing a stable version, and the docs website needs to be updated and
include them.
The only sensible solution here is to pick up the release notes as of
the commit that triggers the release. As the docs cron runs
asynchronously, this means walking down the git history to find the
relevant commit.
> Note: We could probably do away with the asynchronicity at this point.
> It was originally included to cover for the possibility of a release
> failing. If we are releasing commits from the past after they have been
> tested, this should not be an issue anymore. If the docs generation were
> part of the synchronous release step, it would have direct access to the
> correct release notes without having to walk down the git history.
>
> However, I think it is more prudent to keep this change as a future step,
> after we're confident the new release scheme does indeed produce much more
> reliable "stable" releases.
New release process
===================
Just like releases are currently controlled mostly by detecting
changes to the `VERSION` file, the new process will be controlled by
detecting changes to the `LATEST` file. The format of that file will
include both the version string and the corresponding SHA.
Upon detecting a change to the `LATEST` file, CI will run the entire
release process, just like it does now with the VERSION file. The main
differences are:
1. Before running the release step, CI will checkout the commit
specified in the LATEST file. This requires separating the release
step from the build step, which in my opinion is cleaner anyway.
2. The `//:VERSION` Bazel target is replaced by a repository rule
that gets the version to build from an environment variable, with a
default of `0.0.0` to remain consistent with the current `daml-head`
behaviour.
Some of the manual steps will need to be skipped for a snapshot release.
See amended `release/RELEASE.md` in this commit for details.
The main caveat of this approach is that the official release will be a
different binary from the corresponding snapshot. It will have been
built from the same source, but with a different version string. This is
somewhat mitigated by Bazel caching, meaning any build step that does
not depend on the version string should use the cache and produce
identical results. I do not think this can be avoided when our artifact
includes its own version number.
I must note, though, that while going through the changes required after
removing the `VERSION` file, I have been quite surprised at the sheer number of
things that actually depend on the SDK version number. I believe we should
look into reducing that over time.
CHANGELOG_BEGIN
CHANGELOG_END
We have seen a number of failures recently where the collect_build_data
and notify_user steps failed to fetch their branch commit from GitHub.
Trying to reproduce locally doesn't work (i.e. fetching the same sha
succeeds), so we're currently assuming transient network errors.
This PR adds a retry mechanism as well as a Slack message to help me
keep tabs on the issue.
CHANGELOG_BEGIN
CHANGELOG_END
* Update CI nix version
For `--option http2 false` to take effect requires Nix 2.3.2.
CHANGELOG_BEGIN
CHANGELOG_END
* Set option `http2 = false` dev-env nix config
This is less likely to overlook an instance than manually adding
`--option http2 false` to each Nix invocation.
Setting `--option htt2p false` also had no effect on the multi-user Nix
installation on the Linux CI machines due to
```
WARNING: option '--disk_cache' was expanded to from both option '--config linux' (source /nix/store/2xnfb2l39d2b4nxw5vwmqz5hjwhw0caw-daml-bazelrc) and option '--config linux' (source /nix/store/2xnfb2l39d2b4nxw5vwmqz5hjwhw0caw-daml-bazelrc)
```
Co-authored-by: Andreas Herrmann <andreash87@gmx.ch>
The $(Agent.JobStatus) variable unfortunately only tracks the status of
the current _job_, whereas what we really want here is the status of the
whole build. Azure does not have an easy way to do that, but fortunately
we already have logic to determine the current build status in
`collect_build_data`, so here we can just piggyback on that.
CHANGELOG_BEGIN
CHANGELOG_END
PR #4286 introduced new jobs that do not work well when ran against the
master branch, rather than as part of a PR. This hopefully fixes that,
though it's hard to test for obvious reasons.
CHANGELOG_BEGIN
CHANGELOG_END
One of the outputs of our brainstorming about how to make CI better was
that it is annoying to have to "babysit" pull requests. This PR attempts
to introduce a notification mechanism by which Azure will notify people
on Slack when a build finishes, so they know they need to go and rerun
or merge the corresponding PR.
This commit also changes the existing $Slack.URL variable to
$Slack.team-daml, to make more explicit where the Slack message is being
sent to (Slack works with one token per destination channel). Both
$Slack.URL and $Slack.team-daml are currently defined as the same token
in Azure.
CHANGELOG_BEGIN
CHANGELOG_END
Azure builds on the merge commit provided by GitHub as
refs/pull/<pr-number>/merge. Either GitHub recently changed to clear out
such commits faster than they used to, or Azure recently changed to
cache the resulting commit sha rather than go through the indirection
again.
Either way, the end result is that, currently, if the other jobs take
"long enough", and `master` has changed in-between the build starting
and the `collect_build_data` step running, the latter will fail to
checkout the commit it is looking for, and the build will irredeemably
fail. The only option is to re-run the entire build (`/azp run` or
rebase/push), which is sort of the entire opposite of the whole reason
for introducing `collect_build_data` in the first place.
This patch aims to address this by not relying on Azure to fetch the
daml repo in the `collect_build_data` job. This is definitely a hack,
but hopefully one that can alleviate the problem for now.
CHANGELOG_BEGIN
CHANGELOG_END
The main goal of this job is to fail when other, required jobs are
canceled. The reason for this is that the communication between Azure
and GitHub does not always work very well, particularly around canceled
jobs, so that when a job gets canceled GitHub does not always know about
it, and furthermore GitHub does not provide a "re-run" button for
canceled jobs.
Thus, this one provides a failed job that doe sdisplay a "re-run"
button, which re-runs all the failed/canceled jobs in the current build.
Therefore, this only needs to detect canceled jobs, not failed ones
(because those will have their own "re-run" button).
Additionally, we recently changed the standard-change label check to not
run on master (as it really only makes sense on PRs), resulting in a
Skipped status instead of Succeeded, which made this job fail.
CHANGELOG_BEGIN
CHANGELOG_END
This commit:
- fixes an issue introduced by e2015e2ec whereby the data was never
saved to GCS, because if the variable _is_ set, then Azure will
substitute the right side of the comparison as well as set the
environment variable, so the two sides will still be equal. The new
approach skips the first character of the env var and removes the dollar
from the (intended) literal expression, ensuring Azure does not
substitute.
- adds output to the failure case because we have recently seen all
commits on master fail that step, regardless of success of other jobs.
CHANGELOG_BEGIN
CHANGELOG_END
There was an issue where multiline commit messages were not encoded at
all, yielding invalid JSON. This should fix it. This commit message is
designed to test it before it gets to master.
$(exit 1)
`exit 1`
end quote: "
invalid json input
* Rename hie-core to ghcide
The name `hie-core` has caused a lot of confusion as to how we relate
to haskell-ide-engine so changing it should hopefully help with that.
I also think that ghcide is still a good name once we hopefully
integrate with haskell-ide-engine more closely.
The name ghcide seems to have a reasonable amount of support on
Twitter https://twitter.com/ndm_haskell/status/1170681262987710464
which is of course the only good way to come up with names.
* Add a readme that points people to the new directory.
* Fix bogus replacements
* Use a proper link
* links are hard
At the moment, if we try to run the build of a release commit after it
has succeeded, the `git tag` step fails. We do not normally try to
rebuild old commits that had succeeded, but sometimes Azure gets
confused when we ask for rerunning specific jobs within a build, and
reruns the whole build.
This creates two (small, bubt annoying) problems:
1. It adds noise to the #team-daml channel as it notifies of the build
failure, and
2. It marks the commit as failed with a red cross on the github list of
commits, which obviously doesn't look great for release commits.
This fixes that. Note that if a release does fail after the `git tag`
step (e.g. some network error between Azure and GitHub), this **does
not** change the necessary steps to remediate, as that situation would
already be broken in the current setup. (Steps to remediate would
essentially boil down to deleting the tag on GitHub before rerunning so
the build can create it again.)
