The reason why those workflows were split in the first place was just to obtain loadtest
results quicker, because the in the separated workflow, only the single loadtest job
needs to finish before the artifacts can be downloaded.
However, the disadvantage of this approach was, that the results were not as easily
accessible as they could be in a single workflow. Additionally, it's possible to depend
on the "prepopulate nix" job for efficiency if the loadtest runs in the main workflow.
The two different PR and Merge jobs were introduced to be able to test the main branch
against the latest release. However, this is now included in the PR job, too, so no need
for the two separate jobs anymore.
The nix actions cache currently leads to repeated "no space left on
devices" errors for jobs in CI.
The prepopulate job is useless without the nix actions cache, so it
will go away at the same time.
The cache key for nix now depends on default.nix and shell.nix in the root folder and all
.patch files in the nix folder. Those may change the output of our nix derivations, so
must be included. At the same time, there is no reason to include the actions/setup-nix
folder. This would only lead to new caches being created every time we update one of the
dependent actions in this file. Finally, we never restore caches with a different id any-
more. There is no point in having the style job fall back to the static cache for example.
The cache keys for cabal can be more explicit: We only have one postgrest.cabal and one
relevant cabal.project file. We were missing the cabal.project.freeze file, though, which
affects the dependencies used, too.
It used to be possible to spin up a nix environment via docker container this way,
but the upstream nixos/nix image has changed and the docker build doesn't succeed
anymore. Since nobody complained about that, we can assume it is not being used
anyway.
cachix-action v14 added a new daemon mode, which pushes new derivations to the store
as soon as they have been built. This replaces the seed cachix workflow nicely by just
pushing from all jobs directly.
This works around https://github.com/NixOS/nixpkgs/issues/286285 to use -split-sections
in a cross-compiling scenario. This will reduce the size of the static executable and also
remove the remaining references to /nix/store/.. reducing closure size dramatically.
This also fixes the docker image blowing up in size since we switched to pkgsStatic.
By passing -split-sections to all dependencies, GHC will link only the
modules we actually use and not the full package for each dependency.
This does neither work on MacOS nor Windows, thus we don't do it for
stack right now.
Stripping unused symbols in CI will further decrease the size of those files.
This makes the static build fail in case any references to the nix store
are left over. Those will increase the closure size of the nix derivation
massively and lead to a huge docker image.
At the same time, those references will not be functional on non-nix systems,
to which the static executable is distributed, anyway.
PR #2358 added a bulk insert to the loadtest. However this broke the regular insert test,
which just returned 400 Bad Request because of a missing PK value since. Adding the new
id column in the payload to the ?columns= argument fixes that.
The recent nixpkgs update gave us a new version of vegeta. This version includes a new
DNS cache features - which unfortunately doesn't play well with unix sockets. Disabling
the DNS cache makes requests succeed again.
Previously this command upgraded to the latest unstable version of nixpkgs,
but this was often broken. Taking the latest stable branch should give
better results.
This partially reverts #1257 / #1272 / 553531711b where the 404 was introduced.
A 406 error is still returned when requesting a single object via accept header.
Returning an error when no rows are changed can be introduced through a different syntax again, see the discussion in #2164.
Fixes#2343
Signed-off-by: Wolfgang Walther <walther@technowledgy.de>
This was introduced in d5b92a433a. Before
this change, the OpenApi output would have <pk/> annotations for views,
too. After this change, they got lost for mode follow-privileges, because
the pks are refined in haskell code, but the request only fetches all
the tables again, but not the view dependencies.
This fix changes follow-privileges to only fetch a list of accessible
tables, which is then used to filter the tables in the schema cache.
Fixes#2356
Signed-off-by: Wolfgang Walther <walther@technowledgy.de>
This fixes a regression introduced in d2719420f4.
The order of relColumns in the schema cache is now consistently matching the order of columns in foreign key definitions.
Resolves#2518
Signed-off-by: Wolfgang Walther <walther@technowledgy.de>
Aggregates and Window functions can't be called as RPCs in a useful way.
Procedures are not supported right now, but might be added later.
Resolves#2101
Signed-off-by: Wolfgang Walther <walther@technowledgy.de>
The PG version is only read by the Connection Worker, which is not used in the case dump-schema. Now, the pg version is read in the schema cache queries directly, avoiding this problem in all cases.
Signed-off-by: Wolfgang Walther <walther@technowledgy.de>
Trigger functions can't be called directly from SQL and can't be called
via the /rpc prefix either - it makes no sense to expose them in the
OpenAPI output. And we don't need to cache them in the schema cache
either.
Best practice would be to keep the trigger functions in a non-exposed
schema anyway.
