HGitaly
HGitaly is Gitaly server for Mercurial.
It implements the subset of the Gitaly gRPC protocol that is relevant for
Mercurial repositories, as well as its own HGitaly protocol, with methods
that are specific to Mercurial.
It comes in two overlapping variants:
-
HGitaly proper is written in Python, using the grpcio
official library.
-
RHGitaly is a high-performance partial implementation written in Rust, and
based on the tonic
gRPC framework.
As of this writing, RHGitaly implements a strict subset of the methods
implemented in HGitaly, but it is possible that some methods would be
implemented in RHGitaly only in the future.
Installation
HGitaly (Python)
In what follows, $PYTHON
is often the Python interpreter in a virtualenv,
but it can be a system-wide one (typical case in containers, strongly
discouraged on user systemes).
-
Install Mercurial with Rust parts (for the exact version, refer to the
requirements file in the Heptapod main repository sources)
$PYTHON -m pip install --no-use-pep517 --global-option --rust Mercurial==6.6.2
-
Install HGitaly itself (check that it does not reinstall Mercurial)
$PYTHON -m pip install hgitaly
RHGitaly
We distribute a self-contained source tarball. It includes the appropriate
hg-core
Rust sources.
-
Fetch the tarball
wget https://download.heptapod.net/rhgitaly/rhgitaly-x.y.z.tgz
-
Fetch and verify the GPG signature
wget https://download.heptapod.net/rhgitaly/rhgitaly-x.y.z.tgz.asc
gpg --verify rhgitaly-x.y.z.tgz.asc
-
Build
tar xzf rhgitaly-x.y.z.tgz
cd rhgitaly-x.y.z/rust
cargo build --locked --release
-
Install wherever you want. Example given for a system-wide installation
sudo install -o root -g root target/release/rhgitaly /usr/local/bin
-
Define a service. Example given for systemd, to be adjusted for your needs.
Make sure in particular that user and all directories exist, with
appropriate permissions.
[Unit]
Description=Heptapod RHGitaly Server
[Service]
User=hgitaly
# HGRCPATH not needed yet but probably will be at some point
Environment=HGRCPATH=/etc/heptapod/heptapod.hgrc
Environment=RHGITALY_LISTEN_URL=unix:/run/heptapod/rhgitaly.socket
Environment=RHGITALY_REPOSITORIES_ROOT=/home/hg/repositories
ExecStartPre=rm -f /run/heptapod/rhgitaly.socket
ExecStart=/user/local/bin/rhgitaly
Restart=on-failure
[Install]
WantedBy=default.target
External executables
HGitaly needs several other programs to be installed and will run them
as separate processes.
By default, it expects to find them on $PATH
, but the actual path to
each executable can be configured.
Tokei
Tokei is a programming languages analysis
tool written in Rust. It is used by the CommitLanguages
method.
Tokei is available in several Linux distributions.
As of this writing, HGitaly supports versions 12.0 and 12.1
Go license-detector
Usually installed as license-detector
, this standalone executable is
part of the go-enry
suite. Its library version is also used by Gitaly.
It is used in the FindLicense method.
Git
HGitaly can make use of some Git commands that do not involve repositories!
This is for example the case of GetPatchID: the
git patch-id
command does not access any repository. Instead it computes any
patch into an identifier.
Mercurial
In forthcoming versions, it is probable that HGitaly and/or RHGitaly will
invoke Mercurial subprocesses.
This is not yet the case as of this writing (HGitaly 1.1 / Heptapod 1.1).
Configuration
HGitaly's configuration is done the standard way in the Mercurial world:
through HGRC files.
In a typical Heptapod installation, these are split into a managed file, for
consistency with other components and another one for edit by the systems
administrator (/etc/gitlab/heptapod.hgrc
in Omnibus/Docker instances).
Many Mercurial tweaks are interpreted simply because HGitaly internally
calls into Mercurial, but HGitaly also gets its own section. Here are the
settings available as of HGitaly 1.1
[hgitaly]
# paths to external executables
tokei-executable = tokei
license-detector-executable = license-detector
git-executable = git
# The number of workers process default value is one plus half the CPU count.
# It can be explicitly set this way:
#workers = 4
# Time to let a worker finish treating its current request, if any, when
# gracefully restarted. Default is high because of backup requests.
worker.graceful-shutdown-timeout-seconds = 300
# Maximum allowed resident size for worker processes (MiB).
# They get gracefully restarted if they cross that threshold
worker.max_rss_mib = 1024
# Interval between memory monitoring of workers (results dumped in logs)
worker.monitoring-interval-seconds = 60
Also heptapod.repositories-root
is used if --repositories-root
is
not passed on the command line.
Operation
Logging
HGitaly is using the standard logging
Python module, and the
loggingmod
Mercurial extension to emit logs from the Mercurial core
and other extensions. Therefore, the logging configuration is done
from the Mercurial configuration, typically from one of the Heptapod
HGRC files.
The general convention is that all logs emitted by hgitaly.service
provide GitLab's correlation_id
in the extra
dict, making it
available in the format string. Here is a minimal example:
[correlation_id=%(correlation_id)s] [%(levelname)s] [%(name)s] %(message)s"
Conversely, the format strings for logs emitted outside of
hgitaly.service
must not use correlation_id
, as subpackages such as
hgitaly.branch
, hgitaly.message
, etc. cannnot provide a
value: it is a hard error to use a format that relies on some
extra if the emitter does not provide it.
To summarize the resulting policy:
- in
hgitaly.service
, all logging must be done through
hgitaly.logging.LoggerAdapter
. Using correlation_id
in the
format is strongly encouraged. - outside of
hgitaly.service
, logging should be self-contained
useful without an obvious link to the calling gRPC method. For
instance a repository inconsistency should be logged at WARNING
level, with a message including the path.
Development
Automated tests and Continuous Integration
How to run the tests
Usually, that would be in a virtualenv, but it's not necessary.
python3 -m pip install -r test-requirements.txt
./run-all-tests
Hint: Check the contents of run-all-tests
, it's just pytest
with
a standard set of options (mostly for coverage, see below).
Unit and Mercurial integration tests
These are the main tests. They lie inside the hgitaly
and hgext3rd.hgitaly
Python packages. The layout follows the style where
each subpackage has its own tests package, to facilitate future refactorings.
The Mercurial integration tests are written with the mercurial-testhelpers
library. Their duty is to assert that HGitaly works as expected and maintains
compatibility with several versions of Mercurial and possibly other
dependencies, such as grpcio.
The implicit assumption with these tests is that the test authors actually
knew what was expected. HGitaly being meant to be a direct replacement, or
rather a translation of Gitaly in Mercurial terms, those expectation are
actually a mix of:
- Design choices, such as mapping rules between branch/topic combinations
and GitLab branches.
- Gitaly documentation and source code.
- sampling of Gitaly responses.
Gitaly comparison tests
If an appropriate Gitaly installation is found, run-all-tests
will also
run the tests from the tests_with_gitaly
package. This happens automatically
from within a HDK workspace.
These are precisely meant for what the Mercurial integration tests can't do:
check that HGitaly responses take the form expected by the various Gitaly
clients, by comparing directly with the reference Gitaly implementation.
The comparisons work by using the conversions to Git provided by
py-heptapod
, which are precisely what HGitaly aims to replace as a mean
to expose Mercurial content to GitLab.
Once there is no ambiguity with what Gitaly clients expect, the correctness
of the implementation, with its various corner cases,
should be left to the Mercurial integration tests.
Test coverage
This project is being developed with a strong test coverage policy, enforced by
CI: without the Gitaly comparison tests, the coverage has to stay at 100%.
This does not mean that a contribution has to meet this goal to be worthwile,
or even considered. Contributors can expect Maintainers to help them
achieving the required 100% coverage mark, especially if they are newcomers.
Of course, Contributors cannot expect Maintainers to go
as far as write missing tests for them, even if that can still happen
for critical urgent issues.
Selected statements can of course be excluded for good reasons, using
# pragma no cover
.
Coverage exclusions depending on the Mercurial version are
provided by the coverage plugin of mercurial-testhelpers.
Unexpected drop of coverage in different Mercurial versions is a powerful
warning system that something not obvious is getting wrong, but the
Gitaly comparison tests are run in CI against a fixed set of
dependencies, hence 100% coverage must be achieved without the Gitaly
comparison tests.
On the other hand, Gitaly comparison tests will warn us when we bump upstream
GitLab if some critical behaviour has changed.
Tests Q&A and development hints
Doesn't the 100% coverage rule without the Gitaly comparison tests mean writing the same tests twice?
In some cases, yes, but it's limited.
For example, the comparison tests
can tell us that the FindAllBranchNames
is actually expected to return
GitLab refs (refs/heads/some-branch
), not GitLab branch names. That can
be settled with a few, very basic, test cases. There is no need to test all
the mapping rules for topics, and even less the various related corner cases
in the comparison tests. These, on the other hand depend strongly on Mercurial
internals, and absolutely have to be fully tested continuously against various
Mercurial versions.
Also, it is possible to deduplicate scenarios that are almost identical in
Mercurial integration tests and Gitaly comparison tests: factorize out the
common code in a helper function made available for both. The question is if
it is worth the effort.
Finally, comparison tests should focus on the fact that Gitaly and HGitaly
results agree, not on what they contain. In the above example,
a comparison for FindAllBranchNames
could simply assert equality of the
returned sets of branch names. This is a bit less cumbersome, and easier
to maintain.
How to reproduce a drop in coverage found by the compat
CI stage?
These are often due to statements being covered by the Gitaly comparison
tests only, leading to 100% coverage in the main
stage, but not in the
compat
stage.
The first thing to do is to run without the Gitaly comparison tests:
SKIP_GITALY_COMPARISON_TESTS=yes ./run-all-tests
(any non empty value in that environment variable, even no
or false
will
trigger the skipping)
In some rare cases, the drop in coverage could be due to an actual change
between Mercurial versions. If that happens, there are good chances that an
actual bug is lurking around.
How to run the tests with coverage of the Gitaly comparison tests
./run-all-tests --cov tests_with_gitaly --cov-report html
The HTML report will be nice if you don't have 100% coverage. To display it,
just do
firefox htmlcov/index.html
By default, the Gitaly comparison tests themselves are not covered, indeed.
This is because run-all-tests
does not know whether they will be skipped for
lack of a Gitaly installation – which would be legitimate.
But they are covered in the CI jobs that launch them, because Gitaly is
assumed to be available. For these, the coverage would tell us that something
was broken, preventing the tests to run.
How to poke into Gitaly protocol?
The Gitaly comparison tests provide exactly a harness for that: take a test,
modify it as needed, insert a pdb
breakpoint, and get going.
The big advantage here is that startup of the Gitaly comparison tests is
almost instantaneous, especially compared with RSpec, wich takes about a minute
to start even a completely trivial test.
Of course that will raise the question whether it'll be useful to make true
tests of these experiments.
When is a Gitaly comparison test required?
Each time there's a need to be sure of what's expected and it can help answer
that question. It doesn't have to do more than that.
When to prefer writing RSpec tests in Heptapod Rails over Gitaly comparison tests in HGitaly?
If you need to make sure that Heptapod Rails, as a Gitaly client, sends
the proper requests, because that can depend on specific dispatch code.
For instance, we are currently still converting to Git on the Rails side.
A source of bugs would be to send Git commit ids to HGitaly.
Apart from that, it is expected to be vastly more efficient to use
Gitaly comparison tests.
The more Heptapod progresses, the less complicated all of this should be.
Updating the Gitaly gRPC protocol
The virtualenv has to be activated
-
pip install -r dev-requirements.txt
-
Copy the new proto
files from a Gitaly checkout with
version matching the wanted GitLab upstream version.
Example in a HDK context:
cp ../gitaly/proto/*.proto protos/ # we dont want the `go` subdir
-
run ./generate-stubs
-
run the tests: ./run-all-tests
-
perform necessary hg add
after close inspection of hg status
Updating the HGitaly specific gRPC protocol
This package defines and implements an additional gRPC protocol, with
gRPC services and methods that are specific to Mercurial, or more generally
Heptapod.
Protocol specification
The sources are proto
files in the protos/
directory, same as for the
Gitaly protocol.
They distinguish themselves by this declaration:
package hgitaly;
Each time a change is made to the protocol, the libraries for all
provided programming languages have to be regenerated and committed, ideally
together with the protocol change.
Python library
It has a special status, being versioned together with the protocol and the
server implementation. It is provided as the hgitaly.stub
package.
The Python stubs are produced by the same script that takes care of Gitaly
proto
files:
./generate-stubs
Ruby library
See the separate documentation
Other languages
A Go library will probably be necessary quite soon for Workhorse or perhaps
Heptapod Shell.
A Rust library would be nice to have