Right now on errors, only the expected and the actual responses are
shown. The actual response sometimes may not have all the information,
and you may have to look at the logs. In this case, request id would be
of great help to get the extra information from the logs.
These aren't suitable e.g. for running in CI since some take far too
long (and an impossibly long-time when running under criterion's normal
bootstrapping sampling regime.
We might try to improve this ourselves:
https://github.com/bos/criterion/issues/218
An initial summary analysis will be in #3530.
* export metadata without nulls, empty arrays
* property tests for 'ReplaceMetadata' using QuickCheck
-> Derive Arbitrary class for 'ReplaceMetadata' dependant types
* reduce property test cases number to 30
QuickCheck generates the `ReplaceMetadata` value really large
for higher number test cases. Encoded JSON for such values is large and
consumes more memory. Thus, CI is giving up while running property
tests.
* circle-ci: Add property tests as saperate job
* add no command mode to tests
* add yaml.v2 to go mod
* remove indirect comment for yaml.v2 dependency
The connection handler in websocket transport was not using the
'UserAuthentication' interface to resolve user info. Fix resolving
user info in websocket transport to use the common
'UserAuthentication' interface
Instead of
'WITH some_alias (SELECT * from some_func()) SELECT <rows> FROM some_alias'
for SQL function queries, Use
'SELECT <rows> FROM some_func() AS some_alias'
The intention was to make this two cases, using a top-level YAML list.
The result was one test with duplicate keys (effectively only running
the second test). This is an error that's now flagged by newer ruamel.
Both tests needed to be "corrected" to pass and need review.
Tested on python 3.5 and 3.7
We make light use of pyenv to set an appropriate python version if
installed. We could easily install a correct version too if we wanted
but that seemed invasive.
The newer ruamel was an annoying upgrade but also offers some
improvements that exposed some test suite issues (fixed later).
We add a new pytest flag `--accept` that will automatically write back
yaml files with updated responses. This makes it much easier and less
error-prone to update test cases when we expect output to change, or
when authoring new tests.
Second we make sure to test that we actually preserve the order of the
selection set when returning results. This is a "SHOULD" part of the
spec but seems pretty important and something that users will rely on.
To support both of the above we use ruamel.yaml which preserves a
certain amount of formatting and comments (so that --accept can work in
a failry ergonomic way), as well as ordering (so that when we write yaml
the order of keys has meaning that's preserved during parsing).
Use ruamel.yaml everywhere for consistency (since both libraries have
different quirks).
Quirks of ruamel.yaml:
- trailing whitespace in multiline strings in yaml files isn't written
back out as we'd like: https://bitbucket.org/ruamel/yaml/issues/47/multiline-strings-being-changed-if-they
- formatting is only sort of preserved; ruamel e.g. normalizes
indentation. Normally the diff is pretty clean though, and you can
always just check in portions of your test file after --accept
fixup
* save permissions, relationships and collections in catalog with 'is_system_defined'
* Use common stanzas in the .cabal file
* Refactor migration code into lib instead of exe
* Add new server test suite that exercises migrations
* Make graphql-engine clean succeed even if the schema does not exist
This fix is a little ugly, but it’s the only simple solution without a
significant refactoring that restructures the relationship between
GraphQL/Validate and GraphQL/Resolve. The ugliness should go away if we
implement something like #2801.
* Separate DB and metadata migrations
* Refactor Migrate.hs to generate list of migrations at compile-time
* Replace ginger with shakespeare to improve performance
* Improve migration log messages
The changes in 0c74839934 adjusted the
format of error logs slightly to omit fields instead of including them
with null values. However, this was rarely triggered by this test
because it only looks at the first log message, but log messages can
sometimes be written out of order. This makes the test order-agnostic.
Although brotli itself is MIT-licensed, the Haskell brotli library that provides bindings to it is GPL-licensed, so we cannot use it unless we get a response on haskell-hvr/brotli#1.
* Fix hpc combine error
* Do not perform ciignore
* xfail test jsonb_has_all
* Bring back ciignore
* Refer jsonb_has_all xfaul to the corresponding issue in graphql-engine-internal
Add brotli library dependencies to the server docker image.
The compression feature introduced in #2751 requires brotli shared libraries at runtime. In original PR, adding them to server packager image was missing.
* allow customizing GraphQL root field names, close#981
* document v2 track_table API in reference
* support customising column field names in GraphQL schema
* [docs] add custom column fields doc in API reference
* add tests
* rename 'ColField' to 'ColumnField'
* embed column's graphql field in 'PGColumnInfo'
-> Value constructor of 'PGCol' is not exposed
-> Using 'parseJSON' to construct 'PGCol' in 'FromJSON' instances
* avoid using 'Maybe TableConfig'
* refactors & 'custom_column_fields' -> 'custom_column_names'
* cli-test: add configuration field in metadata export test
* update expected keys in `FromJSON` instance of `TableMeta`
* use `buildSchemaCacheFor` to update configuration in v2 track_table
* remove 'GraphQLName' type and use 'isValidName' exposed from parser lib
* point graphql-parser-hs library git repo to hasura
* support 'set_table_custom_fields' query API & added docs and tests
This fixes an issue where queries could incorrectly be considered
reusable if a variable was used in two positions: one where it affected
SQL generation and one where it did not.
Update trigger is failing if any json/geometry columns are present in
event payload rows. Use '*<>' operator instead of '<>' to compare the
internal binary representation of rows if '<>' doesn’t work.
* initial raster support
* _st_intersects_geom -> _st_intersects_geom_nband
* add tests
* update docs
* improve docs
As requested by @marionschleifer
* new type for raster values
Suggested by @lexi-lambda
* replace `SEUnsafe "NULL"` with SENull
* use positional arguments in SQL functions
* only allow omitting set of last arguments in functions
* disallow omitting of a non default argument in functions
These changes also add a new type, PGColumnType, between PGColInfo and
PGScalarType, and they process PGRawColumnType values into PGColumnType
values during schema cache generation.
This mostly simplifies the RootFlds type to make it clearer what it’s
used for, but it has the convenient side-effect of preventing some
“impossible” cases using the type system.
* Listens for SIGTERM as the termination signal
* Stops accepting new connections once the signal is received
* Waits for all connections to be drained, before shutting down
* Forcefully kills all pending connections after 30 seconds
Currently this does not send a close message to websocket clients, I'd
like to submit that change as a separate pull request, but at least this
solve my biggest concern which is not getting confirmation for mutations
while restarting the server.
* allow altering type of a column iff session vars are defined in permissions
* use a sum type to define dependency reason
* set jwt expiry test's expiry time to 4 seconds
* derive Data instance for necessary types to simplify 'hasStaticExp'
At the moment we can...
...run tests in isolation, generating coverage report:
$ dev.sh test
You can pass args to pytest as well. e.g. to run a specific test:
$ dev.sh test -k "test_jsonb_has_all"
Launch a postgres container with useful dev defaults, with PostGIS,
cleaning up afterwards:
$ dev.sh postgres
Build and launch graphql-engine in dev mode, connecting with a
`postgres` launched above
$ dev.sh graphql-engine
This seems to resolve the issue locally (and has worked in the past),
but it's not clear what exactly is going on here (in particular, why
this should resolve what looks like a memory leak). It certainly seems
like a GHC issue of some sort.
Closes#2565
query templates is a little known feature that lets you template rql
queries and serve them as rest apis. This is not relevant anymore
given the GraphQL interface and getting rid of it reduces the dev
time when adding features in few subsystems.
This feature has never been used outside hasura's internal projects or
documented or exposed through console and hence can safely be removed.
* fix bug in audience check while verifying JWT
- previously the check was converting the audience type into a string
and then comparing with the conf value. all audience types (as it is a
string or URI) will convert to plain strings
- use the Audience type from the jose library for comparing
* add docs for audience
* add issuer check as well
* docs minor syntax fix
* skip audience check if not given in conf
* minor docs update
* qualify import jose library
Currently, we allow tracking of a table with the same name as an already tracked function, and vice-versa. This causes an issue when querying from GraphQL since it will only query the table and not the function. I've made changes to disallow this by throwing an error.
* Bump node-sass to version ^4.12.0
This avoids sass/node-sass#2632, which causes compilation failures on
Node v12.x.
* Do most of the work in /pg_dump in Haskell instead of shell
The shell version caused problems on non-Linux systems since it relied
on the non-POSIX -i option for sed, which works slightly differently on
BSD and macOS.
This PR builds console static assets into the server docker image at `/srv/console-assets`. When env var `HASURA_GRAPHQL_CONSOLE_ASSETS_DIR=/srv/console-assets` or flag `--console-assets-dir=/srv/console-assets` is set on the server, the files in this directory are served at `/console/assets/*`.
The console html template will have a variable called `cdnAssets: false` when this flag is set and it loads assets from server itself instead of CDN.
The assets are moved to a new bucket with a new naming scheme:
```
graphql-engine-cdn.hasura.io/console/assets/
/common/{}
/versioned/<version/{}
/channel/<channel>/<version>/{}
```
Console served by CLI will still load assets from CDN - will fix that in the next release.
Changes compared to `/v1alpha1/graphql`
* Changed all graphql responses in **/v1/graphql** endpoint to be 200. All graphql clients expect responses to be HTTP 200. Non-200 responses are considered transport layer errors.
* Errors in http and websocket layer are now consistent and have similar structure.
1. Reuses postgres connections during startup which reduces the overhead of opening and closing connections.
2. Faster schema cache building. This is done by fetching all the required data in a single sql statement.
* build schema cache function without db setup
The setup shouldn't happen for sync. The database is already setup by the instance which generated the event. This means that the sync is now faster.
* use SQL loop to drop hdb_views schema views and routines with ordering
This avoids deadlocks when schema is being changed concurrently
* schema sync now only processes the latest event
This becomes useful when a lot of schema change
events happen while we are still processing an
earlier event.
* split stm transactions when snapshotting to make it faster
* mx subs: push to both old and new sinks at the same time
* expose dev APIs through allowed APIs flag
* add types to represent unparsed http gql requests
This will help when we add caching of frequently used ASTs
* query plan caching
* move livequery to execute
* add multiplexed module
* session variable can be customised depending on the context
Previously the value was always "current_setting('hasura.user')"
* get rid of typemap requirement in reusable plan
* subscriptions are multiplexed when possible
* use lazytx for introspection to avoid acquiring a pg connection
* refactor to make execute a completely decoupled module
* don't issue a transaction for a query
* don't use current setting for explained sql
* move postgres related types to a different module
* validate variableValues on postgres before multiplexing subs
* don't user current_setting for queries over ws
* plan_cache is only visible when developer flag is enabled
* introduce 'batch size' when multiplexing subscriptions
* bump stackage to 13.16
* fix schema_stitching test case error code
* store hashes instead of actual responses for subscriptions
* internal api to dump subscriptions state
* remove PlanCache from SchemaCacheRef
* allow live query options to be configured on server startup
* capture metrics for multiplexed subscriptions
* more metrics captured for multiplexed subs
* switch to tvar based hashmap for faster snapshotting
* livequery modules do not expose internal details
* fix typo in live query env vars
* switch to hasura's pg-client-hs
* read version from env var at build time (close#1398)
* remove un-used imports, edit makefile
* edit makefile to add new targets and export variables
* only export VERSION in makefile
* read version by executing the script if env var is absent
Examples
1) `
pytest --hge-urls "http://127.0.0.1:8080" --pg-urls "postgresql://admin@127.0.0.1:5432/hge_tests" -vv
`
2) `pytest --hge-urls "http://127.0.0.1:8080" "http://127.0.0.1:8081" --pg-urls "postgresql://admin@127.0.0.1:5432/hge_tests" "postgresql://admin@127.0.0.1:5432/hge_tests2" -vv
`
### Solution and Design
<!-- How is this issue solved/fixed? What is the design? -->
<!-- It's better if we elaborate -->
#### Reducing execution time of tests
- The Schema setup and teardown, which were earlier done per test method, usually takes around 1 sec.
- For mutations, the model has now been changed to only do schema setup and teardown once per test class.
- A data setup and teardown will be done once per test instead (usually takes ~10ms).
- For the test class to get this behaviour, one can can extend the class `DefaultTestMutations`.
- The function `dir()` should be define which returns the location of the configuration folder.
- Inside the configuration folder, there should be
- Files `<conf_dir>/schema_setup.yaml` and `<conf_dir>/schema_teardown.yaml`, which has the metadata query executed during schema setup and teardown respectively
- Files named `<conf_dir>/values_setup.yaml` and `<conf_dir>/values_teardown.yaml`. These files are executed to setup and remove data from the tables respectively.
#### Running Graphql queries on both http and websockets
- Each GraphQL query/mutation is run on the both HTTP and websocket protocols
- Pytests test parameterisation is used to achieve this
- The errors over websockets are slightly different from that on HTTP
- The code takes care of converting the errors in HTTP to errors in websockets
#### Parallel executation of tests.
- The plugin pytest-xdist helps in running tests on parallel workers.
- We are using this plugin to group tests by file and run on different workers.
- Parallel test worker processes operate on separate postgres databases(and separate graphql-engines connected to these databases). Thus tests on one worker will not affect the tests on the other worker.
- With two workers, this decreases execution times by half, as the tests on event triggers usually takes a long time, but does not consume much CPU.
From `alpha-40` we've been using a `WHERE` clause to fetch required rows and generate mutation response. This has a few limitations like the requirement of a primary key/unique constraint. This also returns inconsistent data on `delete` mutation as mentioned in #1794.
Now, we're using `VALUES (..)` (refer [here](https://www.postgresql.org/docs/current/sql-values.html)) expression to form virtual table rows in `SQL` to generate mutation response.
Internal changes:-
- Not to use primary key/unique constraint columns:-
- Revert back to `ConstraintName` from `TableConstraint` in `TableInfo` type
- Remove `tcCols` field in `TableConstraint` type
- Modify `table_info.sql` and `fetchTableMeta` function `SQL`
- A test case to perform `delete` mutation and returning relational objects.
1. Haskel library `pg-client-hs` has been updated to expose a function that helps listen to `postgres` notifications over a `channel` in this [PR](https://github.com/hasura/pg-client-hs/pull/5)
2. The server records an event in a table `hdb_catalog.hdb_cache_update_event` whenever any `/v1/query` (that changes metadata) is requested. A trigger notifies a `cache update` event via `hasura_cache_update` channel
3. The server runs two concurrent threads namely `listener` and `processor`. The `listener` thread listens to events on `hasura_cache_update` channel and pushed into a `Queue`. The `processor` thread fetches events from that `Queue` and processes it. Thus server rebuilds schema cache from database and updates.
If returning field contains nested selections then mutation is performed in two steps
1. Mutation is performed with returning columns of any primary key and unique constraints
2. returning fields are queried on rows returned by selecting from table by filtering with column values returned in Step 1.
Since mutation takes two courses based on selecting relations in returning field, it is hard to maintain sequence of prepared arguments (PrepArg) generated while resolving returning field. So, we're using txtConverter instead of prepare to resolve mutation fields.
