Pulp

This is totally different on the single-content branch which converts the last call to bulk_create() and all of them occur in the transaction. So I think this is maybe NOTABUG. I missed triage so I can only comment now.

bmbouter wrote:

This is totally different on the single-content branch which converts the last call to bulk_create() and all of them occur in the transaction. So I think this is maybe NOTABUG. I missed triage so I can only comment now.

What do you mean by "single-content branch" ?

There has been some confusion with this one and https://pulp.plan.io/issues/4060. 4060 is a problem within a single task (sync) of duplicates existing in a stream at the same time. This problem is related, but distinct. As noted in the description, this problem exists when 2 or more tasks are happening at once. If a unit passes through QueryExisting as a non-dupe, and while it is waiting in the Queue for the save stage, and another task creates that unit, it this one becomes a dupe and will fail when it is bulk_saved.

I think it is important to document these problems separately since they occur for different reasons. Some possible solutions might fix both at once. For instance, with docker sync we solve both problems by implementing our own save stage which does not use bulk writes. This allows us to catch and handle integrity errors as they happen. Adding this stage to pulpcore would solve both problems at the expense of database write performance. Since db performance could be a bottleneck, it is worth considering other options if we have any-- but those options would need to take into account both problems (dupes in stream, parallel tasks).

bmbouter, What makes you think that this cannot happen?

worker 1: query content ABC  (QueryStage)
  not-found

worker 2: query content ABC (QueryStage)
  not-found

worker 1: create content ABC  (CreateStage)
  Inserted

worker 2: create content ABC  (CreateStage)
  IntegrityError

@jortel, I agree it could happen. We saw the same problem when we worked on base_paths. We thought running both the query and the create in a single transaction would solve it but it did not:

https://github.com/pulp/pulp/pull/3380#issuecomment-380477101

Recapping some of the discussion from last week, we do need to handle IntegrityErrors-with-requery support in the ContentUnitSaver stage. This ticket can track that. We also need to handle it in the ArtifactSaver stage too, which I believe is tracked as issue https://pulp.plan.io/issues/4060

Does that sound right ^?

I thought @jortel was going to work on a general solution that could be applied to anywhere we call bulk_create and write a ticket for it.