| Status | Authors | Coach | DRIs | Owning Stage | Created |
|---|---|---|---|---|---|
| ongoing |
@grzesiek
|
@ayufan
@grzesiek
|
@jporter
@cheryl.li
| devops verify | 2021-09-10 |
CI/CD data time decay
Summary
GitLab CI/CD is one of the most data and compute intensive components of GitLab. Since its initial release in 2012, the CI/CD subsystem has evolved significantly. It was integrated into GitLab in September 2015 and has become one of the most beloved CI/CD solutions.
On February 1st, 2021, GitLab.com surpassed 1 billion CI/CD builds, and the number of builds continues to grow exponentially.
GitLab CI/CD has come a long way since the initial release, but the design of the data storage for pipeline builds remains almost the same since 2012. In 2021 we started working on database decomposition and extracting CI/CD data to a separate database. Now we want to improve the architecture of GitLab CI/CD product to enable further scaling.
Goals
Implement a new architecture of CI/CD data storage to enable scaling.
Challenges
There are more than two billion rows describing CI/CD builds in GitLab.com’s database. This data represents a sizable portion of the whole data stored in PostgreSQL database running on GitLab.com.
This volume contributes to significant performance problems, development challenges and is often related to production incidents.
We also expect a significant growth in the number of builds executed on GitLab.com in the upcoming years.
Opportunity
CI/CD data is subject to time-decay because, usually, pipelines that are a few months old are not frequently accessed or are even not relevant anymore. Restricting access to processing pipelines that are older than a few months might help us to move this data out of the primary database, to a different storage, that is more performant and cost effective.
It is already possible to prevent processing builds that have been archived. When a build gets archived it will not be possible to retry it, but we still do keep all the processing metadata in the database, and it consumes resources that are scarce in the primary database.
To improve performance and make it easier to scale CI/CD data storage we might want to follow these three tracks described below.
- Partition CI/CD builds queuing database tables
- Partition CI/CD pipelines database tables
- Reduce the rate of builds metadata table growth
Reduce the rate of builds metadata table growth
Once a build (or a pipeline) gets archived, it is no longer possible to resume pipeline processing in such pipeline. It means that all the metadata, we store in PostgreSQL, that is needed to efficiently and reliably process builds can be safely moved to a different data store.
Storing pipeline processing data is expensive as this kind of CI/CD data represents a significant portion of data stored in CI/CD tables. Once we restrict access to processing archived pipelines, we can move this metadata to a different place - preferably object storage - and make it accessible on demand, when it is really needed again (for example for compliance or auditing purposes).
We need to evaluate whether moving data is the most optimal solution. We might be able to use de-duplication of metadata entries and other normalization strategies to consume less storage while retaining ability to query this dataset. Technical evaluation will be required to find the best solution here.
Epic: Reduce the rate of builds metadata table growth.
Partition CI/CD pipelines database tables
Even if we move CI/CD metadata to a different store, or reduce the rate of metadata growth in a different way, the problem of having billions of rows describing pipelines, builds and artifacts, remains. We still may need to keep reference to the metadata we might store in object storage and we still do need to be able to retrieve this information reliably in bulk (or search through it).
It means that by moving data to object storage we might not be able to reduce the number of rows in CI/CD tables. Moving data to object storage should help with reducing the data size, but not the quantity of entries describing this data. Because of this limitation, we still want to partition CI/CD data to reduce the impact on the database (indices size, auto-vacuum time and frequency).
Our intent here is not to move this data out of our primary database elsewhere. What want to divide very large database tables, that store CI/CD data, into multiple smaller ones, using PostgreSQL partitioning features.
There are a few approaches we can take to partition CI/CD data. A promising one is using list-based partitioning where a partition number is assigned a pipeline, and gets propagated to all resources that are related to this pipeline. We will assign a partition number using a uniform logical partition ID This is very flexible because we can extend this partitioning strategy at will; for example with this strategy we can assign an arbitrary partition number based on multiple partitioning keys, combining time-decay-based partitioning with tenant-based partitioning on the application level if desired.
Partitioning rarely accessed data should also follow the policy defined for builds archival, to make it consistent and reliable.
Epic: Partition CI/CD pipelines database tables.
For more technical details about this topic see pipeline data partitioning design.
Partition CI/CD builds queuing database tables
While working on the CI/CD Scale blueprint, we have introduced a new architecture for queuing CI/CD builds for execution.
This allowed us to significantly improve performance. We still consider the new solution as an intermediate mechanism, needed before we start working on the next iteration. The following iteration that should improve the architecture of builds queuing even more (it might require moving off the PostgreSQL fully or partially).
In the meantime we want to ship another iteration, an intermediate step towards more flexible and reliable solution. We want to partition the new queuing tables, to reduce the impact on the database, to improve reliability and database health.
Partitioning of CI/CD queuing tables does not need to follow the policy defined for builds archival. Instead we should leverage a long-standing policy saying that builds created more 24 hours ago need to be removed from the queue. This business rule is present in the product since the inception of GitLab CI.
Epic: Partition CI/CD builds queuing database tables.
For more technical details about this topic see pipeline data partitioning design.
Principles
All the three tracks we will use to implement CI/CD time decay pattern are associated with some challenges. As we progress with the implementation we will need to solve many problems and devise many implementation details to make this successful.
Below, we documented a few foundational principles to make it easier for everyone to understand the vision described in this architectural blueprint.
Removing pipeline data
While it might be tempting to remove old or archived data from our databases this should be avoided. It is usually not desired to permanently remove user data unless consent is given to do so. We can, however, move data to a different data store, like object storage.
Archived data can still be needed sometimes (for example for compliance or auditing reasons). We want to be able to retrieve this data if needed, as long as permanent removal has not been requested or approved by a user.
Accessing pipeline data in the UI
Implementing CI/CD data time-decay through partitioning might be challenging when we still want to make it possible for users to access data stored in many partitions.
We want to retain simplicity of accessing pipeline data in the UI. It will require some backstage changes in how we reference pipeline data from other resources, but we don’t want to make it more difficult for users to find their pipelines in the UI.
We may need to add “Archived” tab on the pipelines / builds list pages, but we should be able to avoid additional steps / clicks when someone wants to view pipeline status or builds associated with a merge request or a deployment.
We also may need to disable search in the “Archived” tab on pipelines / builds list pages.
Accessing pipeline data through the API
We accept the possible necessity of building a separate API endpoint / endpoints needed to access pipeline data through the API.
In the new API users might need to provide a time range in which the data has been created to search through their pipelines / builds. To make it efficient it might be necessary to restrict access to querying data residing in more than two partitions at once. We can do that by supporting time ranges spanning the duration that equals to the builds archival policy.
It is possible to still allow users to use the old API to access archived pipelines data, although a user provided partition identifier may be required.
Iterations
All three tracks can be worked on in parallel:
- Reduce the rate of builds metadata table growth.
- Partition CI/CD pipelines database tables.
- Partition CI/CD queuing tables using list partitioning
Status
In progress.
Timeline
- 2021-01-21: Parent CI Scaling blueprint merge request created.
- 2021-04-26: CI Scaling blueprint approved and merged.
- 2021-09-10: CI/CD data time decay blueprint discussions started.
- 2022-01-07: CI/CD data time decay blueprint merged.
- 2022-02-01: Blueprint updated with new content and links to epics.
- 2022-02-08: Pipeline partitioning PoC merge request started.
- 2022-02-23: Pipeline partitioning PoC successful
- 2022-03-07: A way to attach an existing table as a partition found and proven.
- 2022-03-23: Pipeline partitioning design Google Doc (GitLab internal) started:
https://docs.google.com/document/d/1ARdoTZDy4qLGf6Z1GIHh83-stG_ZLpqsibjKr_OXMgc. - 2022-03-29: Pipeline partitioning PoC concluded.
- 2022-04-15: Partitioned pipeline data associations PoC shipped.
- 2022-04-30: Additional benchmarking started to evaluate impact.
- 2022-06-31: Pipeline partitioning design document merge request merged.
- 2022-09-01: Engineering effort started to implement partitioning.
- 2022-11-01: The fastest growing CI table partitioned:
ci_builds_metadata.