-
- Downloads
[FIX] registry: do not signal a phantom cache clear on load
During loading, the registry clears all `ormcache` data multiple times, in order to ensure consistency with the newly loaded module data. Since 083c70bb, this was done by calling `self.clear_caches()`, with the side-effect of signalling to all other worker processes that the cache *needs* to be invalidated, which is actually untrue. If the other workers have any reason to reload their own registries, they will also clear their own cache in the process - there is no need to forcefully invalidate it globally. One could think that combining the pre-fork mode with the `-d <db>` parameter would mitigate this issue, by making all workers inherit from a fully loaded registry, In reality it doesn't work, because they also inherit from the `cache_invalidated=True` flag, that was never cleared in the master process. So despite having a fully loaded registry, the newly forked workers will signal a cache invalidation upon serving their first request. Further, in a multi-tenant setup with large numbers of databases, registries may be recycled and loaded much more frequently than new workers are starting, due to the limited registry LRU, amplifying this effect a bit. ~~ This patch directly clears the cache LRU without going through `clear_cache()`, avoiding setting the `cache_invalidated` flag of the registry, and thus not signalling to other workers. This is similar to what was being done before 083c70bb, where the LRU was dropped like all other lazy properties. X-original-commit: 87aef4e3
Please register or sign in to comment