SQL: Fix metadata handling in to_semver_no_prerelease() function

[Turbo87]

Resolves #3882

[Turbo87]

unfortunately diesel does not reexport the version that they depend on, so we have to add this here ourselves

[Turbo87]

the numeric feature is necessary because the semver_triple returned by to_semver_no_prerelease() is based on the Numeric SQL type.

[pietroalbini]

While the changes here look correct, my main worry is the performance of the already slow reverse dependencies endpoint. Trying to run the old and the new function in a local database dump shows a 16% slowdown:

cratesio-prod=# explain (analyze, buffers) select num from versions;
                                                   QUERY PLAN                                                   
----------------------------------------------------------------------------------------------------------------
 Seq Scan on versions  (cost=0.00..9824.83 rows=367883 width=6) (actual time=0.053..51.207 rows=367883 loops=1)
   Buffers: shared hit=128 read=6018
 Planning time: 0.434 ms
 Execution time: 67.512 ms
(4 rows)

cratesio-prod=# explain (analyze, buffers) select to_semver_no_prerelease(num) from versions;
                                                     QUERY PLAN                                                      
---------------------------------------------------------------------------------------------------------------------
 Seq Scan on versions  (cost=0.00..101795.58 rows=367883 width=32) (actual time=0.241..1983.644 rows=367883 loops=1)
   Buffers: shared hit=166 read=5986
 Planning time: 0.419 ms
 Execution time: 2000.023 ms
(4 rows)

cratesio-prod=# explain (analyze, buffers) select to_semver_no_prerelease_new(num) from versions;
                                                     QUERY PLAN                                                      
---------------------------------------------------------------------------------------------------------------------
 Seq Scan on versions  (cost=0.00..101795.58 rows=367883 width=32) (actual time=0.247..2382.449 rows=367883 loops=1)
   Buffers: shared hit=192 read=5954
 Planning time: 0.198 ms
 Execution time: 2399.200 ms
(4 rows)

cratesio-prod=# 

I'd like to hear @jtgeibel's opinion here too.

[Turbo87]

we could consider using https://www.postgresql.org/docs/current/ddl-generated-columns.html to have a generated (and stored) version_no_prerelease column in the versions table that the rev deps query can use. that way we only run the function when we insert a new version, but not all the time in the rev deps query.

the only downside of this is that generated columns were only introduced in Postgres 12 and apparently we're still using v11 in production.

I think that a 16% slowdown for a correctness bug is worth it though, and with the suggestion above we should have a viable way forward to make it much faster than before too.

[Turbo87]

in our weekly team meeting today we decided to delay this bugfix until the performance concerns are addressed, which likely requires us to resolve #3480 first.

[bors]

☔ The latest upstream changes (presumably #3958) made this pull request unmergeable. Please resolve the merge conflicts.

[Turbo87]

closing in favor of #4022