Data plane integration testing suite #268

jhamon · 2024-01-10T06:47:39Z

Problem

Want to add a much more comprehensive integration testing suite against the data plane. This is a large undertaking because of the amount of variables involved, but this is a start.

Some variables to consider:

REST vs GRPC
pod vs serverless
environment/cloud/region
freshness timeout
index metric, since hybrid search only makes sense for dotproduct

Solution

Conceptually, I want every test run on github actions to target one combination of the factors above. This means some things which might otherwise be handled through test parameterization in a big monolithic test suite that is run with a single command are now toggled through environment variables. This makes it very easy to setup parallel runs with different configurations, but at the cost of some additional complexity. There's currently no single command to run every combo from my local machine in one go.

To run these locally right now, I might have an incantation like this:

PINECONE_API_KEY='nana-nana-nana-nana-batman' \
  METRIC='cosine' \
  USE_GRPC='false' \
  SPEC='{"serverless": {"region": "us-west-2", "cloud": "aws"}}' \
  poetry run pytest tests/integration/data/ -s -vv

In CI, I shove this ugliness into a reusable github action test-data-plane that allow tests to easily be run using different configurations for target environment/cloud/region, REST vs GRPC, and metric type (which is relevant because hybrid search / sparse vector stuff only makes sense in the context of indexes with metric=dotproduct).
Since waiting on data freshness is another bottleneck in the execution, I try to seed some data upfront and use it for all my query/fetch testing.
By using the exact same tests via GRPC and REST for the first time, I'm discovering a lot of small details that don't match. So I'm making adjustments as I go to bring them into alignment, at least on the happy path. Error handling is a complex beast and will need more iteration in the future.
- Migrated some logic from IndexGRPC in to another vector factory class to allow IndexGRPC#upsert to accept Vector or VectorGRPC args. This greatly simplifies test reuse and will be a nice UX improvement for users also because it simplifies docs.
- Refactored a bit in the GRPC "unit" tests which were and still are a disaster. I found myself crawling through those after refactoring the IndexGRPC class to use the new VectoryFactoryGRPC class and triggering a few failures due to missed a few edge cases expressed through the unit tests. I split things up into smaller files and transitioned from using self.<field> everywhere to using pytest fixtures instead. It still sucks, but readability maybe slightly improved.
- An example of something that changed (for alignment reasons) is that in the past an empty fetch result would be None via GRPC or FetchResult(vectors=[]) via REST. The simplicity of None was nice, but now that there is read_unit information to display even when there are no results, the FetchResponse return value makes more sense.
Implemented a vector factory for GRPC that is very similar to the one I created a while ago for the REST code path. It's much much easier to test edge cases in that class alone than in a more integrationy unit test that requires me to mock/patch a bunch of stuff that would otherwise fire off network calls.

To save resources/time while initially building out these tests, for now I'm only running with serverless indexes. But it should be easy to add other configurations soon.

Some possible bugs/questions to follow up:

Seems like $ne and $nin aren't working as documented
Freshness seems much different on dotproduct indexes. Not sure if this is expected.
Need to rename readUnits to read_units before shipping 3.0

High-level coverage so far

Type of Change

Bug fix (non-breaking change which fixes an issue)
New feature (non-breaking change which adds functionality)
Breaking change (fix or feature that would cause existing functionality to not work as expected)
This change requires a documentation update
Infrastructure change (CI configs, etc)
Non-code change (docs, etc)

austin-denoble

Sorry for all the nitpicks and questions, you can obviously ignore and land as-is. This is fantastic, thanks for expanding our coverage, streamlining the integration flows, and packing in additional GRPC quality of life improvements. Also hot on the heels of all the dependency coverage work, wow! 🚢

austin-denoble · 2024-01-12T19:10:08Z

.github/actions/test-data-plane/action.yaml

+outputs:
+  index_name:
+    description: 'The name of the index, including randomized suffix'
+    value: ${{ steps.create-index.outputs.index_name }}


Is this referring to a different create-index step that's present in this action, or is it assuming it's been run prior? It's working properly so I'm just going to assume the scope for steps.<whatever> is global to the whole workflow. 😅

I think this outputs declaration is just accidental copy/pasta from the create-index action I made earlier this week for dependency testing purposes so it can be removed. If I wanted to use it I could by having my test suite write the name of the index it creates to the GITHUB_OUTPUT location (see here in script used by create-index action) but in this case it's not needed since the test creates and deletes the index.

austin-denoble · 2024-01-12T19:12:48Z

.github/workflows/testing-integration.yaml

+          # - euclidean
+          # - dotproduct


Assuming you want to keep these commented out for now, but if not.

Yeah, leaving them out while still fleshing out the tests and iterating heavily. Dotproduct is having some kind of performance issue at the moment the db team are looking into.

austin-denoble · 2024-01-12T19:25:29Z

.github/workflows/testing-dependency.yaml

+          - 3.9
+          - '3.10'
+          - 3.11


Everything but 3.8 is commented out in dependency-matrix-grpc, is that intentional?

Eh, not exactly. Mainly just trying to save some time and resources while iterating. I should bring them back on the grpc job.

austin-denoble · 2024-01-12T19:29:49Z

pinecone/grpc/vector_factory_grpc.py

+    SparseValues as NonGRPCSparseValues
+)
+
+class VectorFactoryGRPC:


Thanks for all the work getting the grpc side of things a bit more aligned with the non-grpc paths. 🙏

austin-denoble · 2024-01-12T19:34:44Z

pinecone/grpc/vector_factory_grpc.py

+    def build(item: Union[GRPCVector, NonGRPCVector, Tuple, Dict]) -> GRPCVector:
+        if isinstance(item, GRPCVector):
+            return item
+        elif isinstance(item, NonGRPCVector):


I don't think it would, but would the VectorFactory benefit from having the flexibility of GRPCVector being added to the union in build?

Perhaps in some circumstances, but it's probably a lot less likely to happen than people trying to pass the NonGRPCVector object into into the GRPC-flavored client. The GRPC-flavored version of this object is annoying because you have to convert metadata dicts yourself into Struct before passing the kwarg param. So the ergonomics leave a lot to be desired.

austin-denoble · 2024-01-12T19:37:20Z

tests/integration/data/conftest.py

+    return random_string(10)
+
+@pytest.fixture(scope='session')
+def idx(client, index_name, index_host):


nitpick: Any specific reason for using idx?

austin-denoble · 2024-01-12T19:39:26Z

tests/integration/helpers/helpers.py

+            total_time += delta_t
+            time.sleep(delta_t)
+
+def poll_fetch_for_ids_in_namespace(idx, ids, namespace):


👍 Lots of goodies for improving thinking around how to streamline the integration flows on typscript side, thank you again.

austin-denoble · 2024-01-12T19:40:09Z

tests/unit/utils/test_convert_to_list.py

@@ -1,5 +1,6 @@
 import pytest
 from pinecone.utils import convert_to_list
+from pinecone import SparseValues


nitpick: Might not be needed.

austin-denoble · 2024-01-12T19:42:05Z

tests/unit_grpc/conftest.py

+def md1():
+    return {"genre": "action", "year": 2021}
+
+
+@pytest.fixture
+def md2():


nitpick: More descriptive metadata1() and metadata2()? Very nitpicky, but since these fixtures are used across tests in other files might help readability.

I find everything about these tests loathsome even after some refactoring. 💀

They were previously using attributes on a test class called self.md1, etc, so I kept the names the same for now just to simplify the refactor.

austin-denoble · 2024-01-12T19:43:39Z

tests/unit_grpc/test_grpc_index_describe_index_stats.py

+        self.config = Config(api_key="test-api-key", host="foo")
+        self.index = GRPCIndex(config=self.config, index_name="example-name", _endpoint_override="test-endpoint")
+
+    def test_describeIndexStats_callWithoutFilter_CalledWithoutFilter(self, mocker):


nitpick: I was going to ask about the duplicative seeming naming for tests in the grpc suite, but it seems to be across the board in some of these files.

jhamon added 18 commits January 10, 2024 01:47

WIP on CI

1fc6d47

WIP

29f5221

WIP

9e22fa8

WIP

9e2c55c

Reorg test config

ade89fb

Add missing annotation declaring fixture

3fcafb2

Adjust fixtures

c66a8ba

Adjust fixtures

c336400

Refactor test config

018b57c

Run all

434ff04

Fix dotproduct tests

3c364ae

Introduce vector factory in IndexGRPC

0189031

Break up grpc unit tests into smaller files

f80b60e

Adjust config

679bbf9

Extend timeout

677cff7

Add fetch integration tests

a8a725d

Fix typo

3264c08

Add query tests

1fc8768

jhamon changed the title ~~Data plane integration tests~~ Data plane integration testing Jan 12, 2024

jhamon changed the title ~~Data plane integration testing~~ Data plane integration testing suite Jan 12, 2024

jhamon requested review from austin-denoble, haruska and rohanshah18 January 12, 2024 10:17

jhamon marked this pull request as ready for review January 12, 2024 10:20

austin-denoble approved these changes Jan 12, 2024

View reviewed changes

Fix GRPC tests

e63047d

jhamon merged commit 93a04ca into spruce Jan 12, 2024
64 checks passed

jhamon deleted the jhamon/data-plane-testing branch January 12, 2024 20:26

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Data plane integration testing suite #268

Data plane integration testing suite #268

jhamon commented Jan 10, 2024 •

edited

Loading

austin-denoble left a comment

austin-denoble Jan 12, 2024

jhamon Jan 12, 2024

austin-denoble Jan 12, 2024

jhamon Jan 12, 2024

austin-denoble Jan 12, 2024

jhamon Jan 12, 2024

austin-denoble Jan 12, 2024

austin-denoble Jan 12, 2024

jhamon Jan 12, 2024

austin-denoble Jan 12, 2024

austin-denoble Jan 12, 2024

austin-denoble Jan 12, 2024

austin-denoble Jan 12, 2024

jhamon Jan 12, 2024

austin-denoble Jan 12, 2024

Data plane integration testing suite #268

Data plane integration testing suite #268

Conversation

jhamon commented Jan 10, 2024 • edited Loading

Problem

Solution

Some possible bugs/questions to follow up:

High-level coverage so far

Type of Change

austin-denoble left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jhamon commented Jan 10, 2024 •

edited

Loading