Query plan caching was introduced by - I believe - hasura/graphql-engine#1934 in order to reduce the query response latency. During the development of PDV in hasura/graphql-engine#4111, it was found out that the new architecture (for which query plan caching wasn't implemented) performed comparably to the pre-PDV architecture with caching. Hence, it was decided to leave query plan caching until some day in the future when it was deemed necessary.
Well, we're in the future now, and there still isn't a convincing argument for query plan caching. So the time has come to remove some references to query plan caching from the codebase. For the most part, any code being removed would probably not be very well suited to the post-PDV architecture of query execution, so arguably not much is lost.
Apart from simplifying the code, this PR will contribute towards making the GraphQL schema generation more modular, testable, and easier to profile. I'd like to eventually work towards a situation in which it's easy to generate a GraphQL schema parser *in isolation*, without being connected to a database, and then parse a GraphQL query *in isolation*, without even listening any HTTP port. It is important that both of these operations can be examined in detail, and in isolation, since they are two major performance bottlenecks, as well as phases where many important upcoming features hook into.
Implementation
The following have been removed:
- The entirety of `server/src-lib/Hasura/GraphQL/Execute/Plan.hs`
- The core phases of query parsing and execution no longer have any references to query plan caching. Note that this is not to be confused with query *response* caching, which is not affected by this PR. This includes removal of the types:
- - `Opaque`, which is replaced by a tuple. Note that the old implementation was broken and did not adequately hide the constructors.
- - `QueryReusability` (and the `markNotReusable` method). Notably, the implementation of the `ParseT` monad now consists of two, rather than three, monad transformers.
- Cache-related tests (in `server/src-test/Hasura/CacheBoundedSpec.hs`) have been removed .
- References to query plan caching in the documentation.
- The `planCacheOptions` in the `TenantConfig` type class was removed. However, during parsing, unrecognized fields in the YAML config get ignored, so this does not cause a breaking change. (Confirmed manually, as well as in consultation with @sordina.)
- The metrics no longer send cache hit/miss messages.
There are a few places in which one can still find references to query plan caching:
- We still accept the `--query-plan-cache-size` command-line option for backwards compatibility. The `HASURA_QUERY_PLAN_CACHE_SIZE` environment variable is not read.
https://github.com/hasura/graphql-engine-mono/pull/1815
GitOrigin-RevId: 17d92b254ec093c62a7dfeec478658ede0813eb7
These must have gotten messed up during a refactor. As a consequence
almost all samples received so far fall into the single erroneous 0 to
1K seconds (originally supposed to be 1ms?) bucket.
I also re-thought what the numbers should be, but these are still
arbitrary and might want adjusting in the future.
We upload a set of accumulating timers and counters to track service
time for different types of operations, across several dimensions (e.g.
did we hit the plan cache, was a remote involved, etc.)
Also...
Standardize on DiffTime as a standard duration type, and try to use it
consistently.
See discussion here:
https://github.com/hasura/graphql-engine/pull/3584#pullrequestreview-340679369
It should be possible to overwrite that module so the new threadDelay
sticks per the pattern in #3705 blocked on #3558
Rename the Control.Concurrent.Extended.threadDelay to `sleep` since a
naive use with a literal argument would be very bad!
We catch a bug in 'computeTimeDiff'.
Add convenient 'Read' instances to the time unit utility types. Make
'Second' a newtype to support this.
