Cloud native and the adoption of Kubernetes has been recognised by GitLab to be one of the top two biggest tailwinds that are helping us grow faster as a company behind the project.
This effort is described in a more details in the infrastructure team handbook.
Traditional job logs depend a lot on availability of a local shared storage.
Every time a GitLab Runner sends a new partial build output, we write this output to a file on a disk. This is simple, but this mechanism depends on shared local storage - the same file needs to be available on every GitLab web node machine, because GitLab Runner might connect to a different one every time it performs an API request. Sidekiq also needs access to the file because when a job is complete, a trace file contents will be sent to the object store.
New architecture writes data to Redis instead of writing build logs into a file.
In order to make this performant and resilient enough, we implemented a chunked I/O mechanism - we store data in Redis in chunks, and migrate them to an object store once we reach a desired chunk size.
Simplified sequence diagram is available below.
In 2017, we experienced serious problems of scaling our NFS infrastructure. We even tried to replace NFS with CephFS - unsuccessfully.
Since that time it has become apparent that the cost of operations and maintenance of a NFS cluster is significant and that if we ever decide to migrate to Kubernetes we need to decouple GitLab from a shared local storage and NFS.
- NFS might be a single point of failure
- NFS can only be reliably scaled vertically
- Moving to Kubernetes means increasing the number of mount points by an order of magnitude
- NFS depends on extremely reliable network which can be difficult to provide in Kubernetes environment
- Storing customer data on NFS involves additional security risks
Moving GitLab to Kubernetes without NFS decoupling would result in an explosion of complexity, maintenance cost and enormous, negative impact on availability.
- ✓ Implement the new architecture in way that it does not depend on shared local storage
- ✓ Evaluate performance and edge-cases, iterate to improve the new architecture
- ✓ Design cloud native build logs correctness verification mechanisms
- ✓ Build observability mechanisms around performance and correctness
- Rollout the feature into production environment incrementally
The work needed to make the new architecture production ready and enabled on GitLab.com is being tracked in Cloud Native Build Logs on GitLab.com epic.
Enabling this feature on GitLab.com is a subtask of making the new architecture generally available for everyone.
|Architecture Evolution Coach||Gerardo Lopez-Fernandez|
|Engineering Leader||Darby Frey|
|Domain Expert||Kamil Trzciński|
|Domain Expert||Sean McGivern|