e43a5e4794
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/4309 GitOrigin-RevId: 899b1aa3674654f5c4bdf94dcd45009570b34644 |
||
---|---|---|
.. | ||
Harness | ||
Test | ||
README.md | ||
Spec.hs | ||
SpecHook.hs |
tests-hspec
Graphql-engine integration tests written in Haskell using the hspec testing framework.
For motivation, rationale, and more, see the test suite rfc.
Table of Contents
- tests-hspec
Running the test suite
-
To run the Haskell integration test suite, we'll first need to start the backends:
docker-compose up
This will start up Postgres, SQL Server, Citus and MariaDB.
Note
: on ARM64 architecture we'll need additional steps in order to test mssql properly.
Preparation
- Switch the docker image in
docker-compose/sqlserver/Dockerfile
toazure-sql-edge
:
- FROM mcr.microsoft.com/mssql/server:2019-latest@sha256:a098c9ff6fbb8e1c9608ad7511fa42dba8d22e0d50b48302761717840ccc26af + FROM mcr.microsoft.com/azure-sql-edge
- Install
sqlcmd
locally. On MacOS, this can be done with brew:brew install mssql-tools
.
Start the backends
-
Run
docker-compose up
-
Initialize the SQL Server database
(cd docker-compose/sqlserver && bash run-init.sh 65003)
- Switch the docker image in
-
Once the containers are up, you can run the test suite via
cabal run tests-hspec
You can also further refine which tests to run using the
-m
flag:cabal run tests-hspec -- -m "SQLServer"
For additional information, consult the help section:
cabal run tests-hspec -- --help
-
The local databases persist even after shutting down docker-compose. If this is undesirable, delete the databases using the following command:
docker-compose down --volumes
Enabling logging
In order to enable logging, you have to manually edit the engineLogLevel
term
in Harness/Constants.hs.
This pairs well with running a single test via the -m
flag (see the section
above).
Test suite structure
Harness
Modules under the Harness/ namespace provide the infrastructure and supporting code for writing and running the tests. It includes quasiquoters, interacting with backends, interfacing with HTTP, constants, and so on.
Supporting code should be added under the Harness.*
namespace instead of
added ad-hoc in test specs, to improve readability and reuse.
Test
Modules under the Test/ namespace define integration test specifications for various features and backends.
Adding a new test
The module Test.HelloWorldSpec contains a starting point which can be built upon.
To create a new test:
- Create a new module based on
Test.HelloWorldSpec
- Specify each relevant context on which the tests will run in
spec
- Specify the tests and expectations in
tests
When creating a new test, make sure that:
- The module name is suffixed by the word
Spec
- The module exports the entry point
spec :: SpecWith State
- The module is listed in the cabal file under
other-modules
(1) and (2) are required for hspec-discover
to find and run the test.
Specifying contexts
We often want to run the same tests several times with slightly different configuration. Most commonly, we want to assert that a given behaviour works consistently across different backends.
Harness.Test.Context defines two functions for
running test trees in terms of a list of Context a
s.
Each Context a
requires:
- a unique
name
, of typeContextName
- a
mkLocalState
action, of typeState -> IO a
- a
setup
action, of type(State, a) -> IO ()
- a
teardown
action, of type(State, a) -> IO ()
- an
options
parameter, which will be threaded through the tests themselves to modify behavior for a particularContext
Of these two functions, whether one wishes to use Harness.Test.Context.run
or
Harness.Test.Context.runWithLocalState
will depend on if their test can be
written in terms of information provided by the global State
type or if it
depends on some additional "local" state.
More often than not, test authors should use Harness.Test.Context.run
, which
is written in terms of Context ()
. This uses ()
for the local test which
does not carry any "useful" state information, and is therefore omitted from
the body of the tests themselves.
In the rare cases where some local state is necessary (either for the test
itself, or as an argument to the teardown
action for some Context
), test
authors should use Harness.Test.Context.runWithLocalState
. This function
takes a type parameter for its local state, which will be provided to both
the teardown
action specified in Context
as well as the body of tests
themselves.
Make local state action
This refers to the function mkLocalState
defined for Context
:
mkLocalState :: State -> IO a
Its return value, IO a
, matches the a
of Context a
: it is the additional
local state that is required throughout the tests, in addition to the global
State
. Some tests, such as tests which check remote relationships, need to keep some state which is local to the context,
but most tests do not need additional state, and define mkLocalState
to be Harness.Test.Context.noLocalState
.
This local state will be pass to the setup
function and the teardown
function.
The teardown
function is responsible to destroy the local state as well, if needed.
Setup action
A setup action is a function of type (State, a) -> IO ()
which is responsible for
creating the environment for the test. It needs to:
- Clear and reconfigure the metadata
- Setup tables and insert values
- Track tables, add relationships, permissions
etc.
These actions can be created by running POST requests against graphql-engine
using Harness.GraphqlEngine.post_
, or by running SQL requests against the
backend using Backend.<backend>.run_
.
Teardown action
The teardown action is another function of type (State, a) -> IO ()
which is
responsible for removing the environment created by the test or setup, so that
other tests can have a "clean slate" with no artifacts. The (State, a)
parameter is constructed from the a
parameter of the Context a
: it is the
local state that is passed throughout the tests.
This action is responsible for freeing acquired resources, and reverting all local modifications: dropping newly created tables, deleting custom functions, removing the changes made to the metadata, and so on.
These actions can be created by running POST requests against graphql-engine
using Harness.GraphqlEngine.post_
, or by running SQL requests against the
backend using Backend.<backend>.run_
.
Writing tests
Test should be written (or reachable from) tests :: SpecWith State
, or tests :: SpecWith (State, Foo)
for tests that use an additional local state.
A typical test will look similar to this:
it "Where id=1" \state ->
shouldReturnYaml
( GraphqlEngine.postGraphql
state
[graphql|
query {
hasura_author(where: {id: {_eq: 1}}) {
name
id
}
}
|]
)
[yaml|
data:
hasura_author:
- name: Author 1
id: 1
|]
it
specifies the name of the testshouldReturnYaml
creates anExpectation
which does the following:- Runs a POST request against graphql-engine which can be specified using the
graphql
quasi-quoter. - Compares the response to an expected result which can be specified using the
yaml
quasi-quoter.
- Runs a POST request against graphql-engine which can be specified using the
Note: these quasi-quoter can also perform string interpolation. See the relevant modules under the Harness.Quoter namespace.
Style guide
Stick to Simple Haskell
This test suite should remain accessible to contributors who are new to Haskell and/or the GraphQL engine codebase. Consider the power-to-weight ratio of features, language extensions or abstractions before you introduce them. For example, try to fully leverage Haskell '98 or 2010 features before making use of more advanced ones.
Write small, atomic, autonomous specs
Small: Keep specs short and succinct wherever possible. Consider reorganising modules that grow much longer than ~200-300 lines of code.
For example: The TestGraphQLQueryBasic*
pytest class was ported to the hspec suite as separate BasicFields
, LimitOffset
, Where
, Ordering
, Directives
and Views
specs.
Atomic: Each spec should test only one feature against the backends (or contexts) that support it. Each module should contain only the context setup and teardown, and the tests themselves. The database schema, test data, and feature expectations should be easy to reason about without navigating to different module.
For example: BasicFieldsSpec.hs
Autonomous: Each test should run independently of other tests, and not be dependent on the results of a previous test. Shared test state, where unavoidable, should be made explicit.
For example: Remote relationship tests explicitly require shared state.
Use the Harness.*
hierarchy for common functions
Avoid functions or types in tests, other than calls to the Harness.*
API.
Any supporting code should be in the Harness.*
hierarchy and apply broadly to the test suites overall.