Prior to this MR there were two GitHub related importers:
* Github::Import: the main importer used for GitHub projects
* Gitlab::GithubImport: importer that's somewhat confusingly used for
importing Gitea projects (apparently they have a compatible API)
This MR renames the Gitea importer to Gitlab::LegacyGithubImport and
introduces a new GitHub importer in the Gitlab::GithubImport namespace.
This new GitHub importer uses Sidekiq for importing multiple resources
in parallel, though it also has the ability to import data sequentially
should this be necessary.
The new code is spread across the following directories:
* lib/gitlab/github_import: this directory contains most of the importer
code such as the classes used for importing resources.
* app/workers/gitlab/github_import: this directory contains the Sidekiq
workers, most of which simply use the code from the directory above.
* app/workers/concerns/gitlab/github_import: this directory provides a
few modules that are included in every GitHub importer worker.
== Stages
The import work is divided into separate stages, with each stage
importing a specific set of data. Stages will schedule the work that
needs to be performed, followed by scheduling a job for the
"AdvanceStageWorker" worker. This worker will periodically check if all
work is completed and schedule the next stage if this is the case. If
work is not yet completed this worker will reschedule itself.
Using this approach we don't have to block threads by calling `sleep()`,
as doing so for large projects could block the thread from doing any
work for many hours.
== Retrying Work
Workers will reschedule themselves whenever necessary. For example,
hitting the GitHub API's rate limit will result in jobs rescheduling
themselves. These jobs are not processed until the rate limit has been
reset.
== User Lookups
Part of the importing process involves looking up user details in the
GitHub API so we can map them to GitLab users. The old importer used
an in-memory cache, but this obviously doesn't work when the work is
spread across different threads.
The new importer uses a Redis cache and makes sure we only perform
API/database calls if absolutely necessary. Frequently used keys are
refreshed, and lookup misses are also cached; removing the need for
performing API/database calls if we know we don't have the data we're
looking for.
== Performance & Models
The new importer in various places uses raw INSERT statements (as
generated by `Gitlab::Database.bulk_insert`) instead of using Rails
models. This allows us to bypass any validations and callbacks,
drastically reducing the number of SQL queries and Gitaly RPC calls
necessary to import projects.
To ensure the code produces valid data the corresponding tests check if
the produced rows are valid according to the model validation rules.
This reduces the time spent waiting for Sidekiq jobs to complete in
JobWaiter, and reduces the sleep interval when trying to acquire the
lease for refreshing authorizations. These changes should reduce the
time spent just waiting for a lock, which we seem to be spending most
time in when running the AuthorizedProjectsWorker.
There were two cases that could be problematic:
1. Because sometimes AuthorizedProjectsWorker would be scheduled in a
transaction it was possible for a job to run/complete before a
COMMIT; resulting in it either producing an error, or producing no
new data.
2. When scheduling jobs the code would not wait until completion. This
could lead to a user creating a project and then immediately trying
to push to it. Usually this will work fine, but given enough load it
might take a few seconds before a user has access.
The first one is problematic, the second one is mostly just annoying
(but annoying enough to warrant a solution).
This commit changes two things to deal with this:
1. Sidekiq scheduling now takes places after a COMMIT, this is ensured
by scheduling using Rails' after_commit hook instead of doing so in
an arbitrary method.
2. When scheduling jobs the calling thread now waits for all jobs to
complete.
Solution 2 requires tracking of job completions. Sidekiq provides a way
to find a job by its ID, but this involves scanning over the entire
queue; something that is very in-efficient for large queues. As such a
more efficient solution is necessary. There are two main Gems that can
do this in a more efficient manner:
* sidekiq-status
* sidekiq_status
No, this is not a joke. Both Gems do a similar thing (but slightly
different), and the only difference in their name is a dash vs an
underscore. Both Gems however provide far more than just checking if a
job has been completed, and both have their problems. sidekiq-status
does not appear to be actively maintained, with the last release being
in 2015. It also has some issues during testing as API calls are not
stubbed in any way. sidekiq_status on the other hand does not appear to
be very popular, and introduces a similar amount of code.
Because of this I opted to write a simple home grown solution. After
all, all we need is storing a job ID somewhere so we can efficiently
look it up; we don't need extra web UIs (as provided by sidekiq-status)
or complex APIs to update progress, etc.
This is where Gitlab::SidekiqStatus comes in handy. This namespace
contains some code used for tracking, removing, and looking up job IDs;
all without having to scan over an entire queue. Data is removed
explicitly, but also expires automatically just in case.
Using this API we can now schedule jobs in a fork-join like manner: we
schedule the jobs in Sidekiq, process them in parallel, then wait for
completion. By using Sidekiq we can leverage all the benefits such as
being able to scale across multiple cores and hosts, retrying failed
jobs, etc.
The one downside is that we need to make sure we can deal with
unexpected increases in job processing timings. To deal with this the
class Gitlab::JobWaiter (used for waiting for jobs to complete) will
only wait a number of seconds (30 by default). Once this timeout is
reached it will simply return.
For GitLab.com almost all AuthorizedProjectWorker jobs complete in
seconds, only very rarely do we spike to job timings of around a minute.
These in turn seem to be the result of external factors (e.g. deploys),
in which case a user is most likely not able to use the system anyway.
In short, this new solution should ensure that jobs are processed
properly and that in almost all cases a user has access to their
resources whenever they need to have access.
