mirror of
https://github.com/opentffoundation/opentf.git
synced 2025-12-19 17:59:05 -05:00
0acefb40c1
Previously we used an API design that suggested that provider schemas were treated as a read-through cache with providers loaded only on demand. However, that's really a fiction: during graph construction the AttachSchemaTransformer would always end up loading all of the schemas up front anyway, and so everything after that point would always be pulling from that cache. To make the situation a little more explicit so that the system is easier to follow -- particularly now that we'll be exposing work like this in OpenTelemetry tracing -- we'll switch to a model where populating the cache is an explicit step and then all downstream calls just assume the cache is already populated. This means that there's now one obvious place where the provider schema lookups are triggered, but we'll also do the real work in the background and thus we can make some progress on the graph construction (CPU bound) concurrently with the schema fetching (I/O bound) so that the schema is more likely to be available at or soon after the start of the execution of AttachSchemaTransformer, which will then block until the schema loading has completed. We have some unit tests that were previously depending on the "load schemas on first request, wherever it happens" behavior and so those are now updated here to either explicitly load the schemas themselves or to rely on a higher-level helper to do it for them, but for normal code and for context tests we expect that the schema loading will always be triggered at the first entry into an exported method of Context, with subsequent calls able to reuse that cache as long as they are still working with the same providers. Signed-off-by: Martin Atkins <mart@degeneration.co.uk>