Error when creating Vector Index for MongoDB vCore

Question

Error when creating Vector Index for MongoDB vCore

Tyler Suard 155

Hello. I uploaded around 150,000 documents to my MongoDB vCore. My searches were returning either terrible results or no results at all, so I decided to recreate the index. That is when I got this error message:

User's image

I can't create an index with fewer documents. Also, if I create an index before inserting documents into the collection, then the index won't contain those documents. How can I fix this issue? MogoDB vCore should be able to handle many documents. In the documentation it mentions creating an index when there are over a million documents.

William Wang 0 Reputation points Microsoft Employee

2024-04-05T02:11:58.31+00:00

Sorry to hear about the issue you are experiencing. Which cluster tier and indexing algorithm are you using?
Tyler Suard 155 Reputation points

2024-04-05T17:17:06.89+00:00

Hello William, thank you for responding. My pricing tier is Standard3. Indexing algorithm: HNSW.
Tyler Suard 155 Reputation points

2024-04-05T17:18:04.2966667+00:00

Thank you for the quick response William. I am using cluster tier Standard 3 and algorithm HNSW.
William Wang 0 Reputation points Microsoft Employee

2024-04-05T17:31:28.4833333+00:00

Which documentation were you referring to when you said "In the documentation it mentions creating an index when there are over a million documents"?
Tyler Suard 155 Reputation points

2024-04-05T17:44:04.66+00:00

Sorry, my mistake, my cluster tier is M30 tier, 2 vCores, 8 GiB RAM, and the algorithm we are using is IVF. The documentation I am referring to is here: https://learn.microsoft.com/en-us/azure/cosmos-db/mongodb/vcore/vector-search
Tyler Suard 155 Reputation points

2024-04-05T17:45:07.5433333+00:00

The documentation talks about indexing more than 1 million documents, so my indexing only 150,000 should not be a problem.
James Codella 10 Reputation points Microsoft Employee

2024-04-05T17:53:12.5+00:00

Hey @Tyler Suard sorry you hit an issue here.

In general if your resource runs out of memory, you can try either scaling up the resource to a higher cluster tier, or scaling down the index parameters.

For more than ~10k ada-002 vectors, I wouldn't recommend anything lower than M40. Also, if you scale up to an M40 or above you'll be able to create an HNSW index. HNSW should offer you better latency, throughput, and accuracy compared to IVF at the scale you have.

Please reach out with any other questions.
Tyler Suard 155 Reputation points

2024-04-05T17:57:13.5533333+00:00

Thank you James. Our database has 800,000 vectors. Will we be able to create a vector index using an M40 machine? What do you mean by scaling down the index parameters? We are using text-embedder-3-small.
Tyler Suard 155 Reputation points

2024-04-05T18:01:27.5733333+00:00

Also, if we update to M40 can we use the database we have already made or do we need to start over from scratch?
James Codella 10 Reputation points Microsoft Employee

2024-04-05T18:03:47.39+00:00

I believe so, I was able to do >1M vectors on an M50 without issue. What size embeddings are you using from text-embedding-3-small? The new embeddings models allow you to set an embedding dimension size, so this is customizable.

By scaling down the index parameters, I mean for HNSW you can reduce m or efConstruction and this will reduce the memory footprint of the index. However, you may lose some accuracy in doing so.
Tyler Suard 155 Reputation points

2024-04-05T18:06:02.4266667+00:00

Ok, we are using the default 1536 dimensions. I will upgrade to M50 now ,thank you.
James Codella 10 Reputation points Microsoft Employee

2024-04-05T18:09:41.61+00:00

You can scale up the resource and maintain your data. No need to start over :)
James Codella 10 Reputation points Microsoft Employee

2024-04-05T18:13:00.8366667+00:00

@Tyler Suard feel free to reach out to me directly at: myFirstName.myLastName@microsoft.com. Would be happy to hear any feedback you might have.
Tyler Suard 155 Reputation points

2024-04-05T18:17:11.4866667+00:00

Thank you James, I appreciate your help.
Tyler Suard 155 Reputation points

2024-04-05T18:44:59.64+00:00
James, I upgraded to an M50 tier. Now I am running the index creation from a python script and get an error.

Here is my code, using pymongo:

vector_db.command( { "createIndexes": "WCS_parts_vector_collection", "indexes": [ { "name": "WCS_parts_vector_index", "key": {"Vector": "cosmosSearch"}, "cosmosSearchOptions": { "kind": "vector-hnsw", "efConstruction": 64, # default value "similarity": "COS", "dimensions": 1536, }, } ], } )

After running it, I get this error:

raise ExecutionTimeout(errmsg, code, response, max_wire_version)

pymongo.errors.ExecutionTimeout: The command being executed was terminated due to a command timeout. This may be due to concurrent transactions. Consider increasing the maxTimeMS on the command., full error: {'ok': 0.0, 'errmsg': 'The command being executed was terminated due to a command timeout. This may be due to concurrent transactions. Consider increasing the maxTimeMS on the command.', 'code': 50, 'codeName': 'ExceededTimeLimit'}
James Codella 10 Reputation points Microsoft Employee

2024-04-05T18:57:35.2933333+00:00

Seems like it's just timing out since it's taking a while to build the index on a large number of vectors. You can increase the Collection operation timeout by collection.maxTimeMS(n)

Try setting it to something large (it's in milliseconds)
Tyler Suard 155 Reputation points

2024-04-05T19:07:18.8033333+00:00

Thank you for your help again James, but I am getting this error when I try to do what you recommended:
Tyler Suard 155 Reputation points

2024-04-05T19:12:13.2233333+00:00
Ok I think I figured it out,

vector_db.command(create index params in here).maxTimeMS(1000000000000000)
Tyler Suard 155 Reputation points

2024-04-05T19:12:51.2233333+00:00

No, actually even after that I still got the timeout error. What am I doing wrong?
Tyler Suard 155 Reputation points

2024-04-05T19:21:41.33+00:00

I can't figure out where to execute the maxTimeMS command, even though I am looking everywhere online.

James Codella 10 Microsoft Employee

vector_db.command(
    {
        "createIndexes": "WCS_parts_vector_collection",
        "indexes": [
            {
                "name": "WCS_parts_vector_index",
                "key": {"Vector": "cosmosSearch"},
                "cosmosSearchOptions": {
                    "kind": "vector-hnsw",
                    "efConstruction": 64,  # default value
                    "similarity": "COS",
                    "dimensions": 1536,
                },
            }
        ],
    }, maxTimeMS=100000000
)

Hmm, sorry for that. Try this, I just verified it should work.

We can update the documentation to provide clearer guidance on how to address the timeout.

1 answer

Your answer

William Wang 0 Reputation points Microsoft Employee

2024-04-05T02:11:58.31+00:00

Sorry to hear about the issue you are experiencing. Which cluster tier and indexing algorithm are you using?
Tyler Suard 155 Reputation points

2024-04-05T17:17:06.89+00:00

Hello William, thank you for responding. My pricing tier is Standard3. Indexing algorithm: HNSW.
Tyler Suard 155 Reputation points

2024-04-05T17:18:04.2966667+00:00

Thank you for the quick response William. I am using cluster tier Standard 3 and algorithm HNSW.
William Wang 0 Reputation points Microsoft Employee

2024-04-05T17:31:28.4833333+00:00

Which documentation were you referring to when you said "In the documentation it mentions creating an index when there are over a million documents"?
Tyler Suard 155 Reputation points

2024-04-05T17:44:04.66+00:00

Sorry, my mistake, my cluster tier is M30 tier, 2 vCores, 8 GiB RAM, and the algorithm we are using is IVF. The documentation I am referring to is here: https://learn.microsoft.com/en-us/azure/cosmos-db/mongodb/vcore/vector-search
Tyler Suard 155 Reputation points

2024-04-05T17:45:07.5433333+00:00

The documentation talks about indexing more than 1 million documents, so my indexing only 150,000 should not be a problem.
James Codella 10 Reputation points Microsoft Employee

2024-04-05T17:53:12.5+00:00

Hey @Tyler Suard sorry you hit an issue here.

In general if your resource runs out of memory, you can try either scaling up the resource to a higher cluster tier, or scaling down the index parameters.

For more than ~10k ada-002 vectors, I wouldn't recommend anything lower than M40. Also, if you scale up to an M40 or above you'll be able to create an HNSW index. HNSW should offer you better latency, throughput, and accuracy compared to IVF at the scale you have.

Please reach out with any other questions.
Tyler Suard 155 Reputation points

2024-04-05T17:57:13.5533333+00:00

Thank you James. Our database has 800,000 vectors. Will we be able to create a vector index using an M40 machine? What do you mean by scaling down the index parameters? We are using text-embedder-3-small.
Tyler Suard 155 Reputation points

2024-04-05T18:01:27.5733333+00:00

Also, if we update to M40 can we use the database we have already made or do we need to start over from scratch?
James Codella 10 Reputation points Microsoft Employee

2024-04-05T18:03:47.39+00:00

I believe so, I was able to do >1M vectors on an M50 without issue. What size embeddings are you using from text-embedding-3-small? The new embeddings models allow you to set an embedding dimension size, so this is customizable.

By scaling down the index parameters, I mean for HNSW you can reduce m or efConstruction and this will reduce the memory footprint of the index. However, you may lose some accuracy in doing so.
Tyler Suard 155 Reputation points

2024-04-05T18:06:02.4266667+00:00

Ok, we are using the default 1536 dimensions. I will upgrade to M50 now ,thank you.
James Codella 10 Reputation points Microsoft Employee

2024-04-05T18:09:41.61+00:00

You can scale up the resource and maintain your data. No need to start over :)
James Codella 10 Reputation points Microsoft Employee

2024-04-05T18:13:00.8366667+00:00

@Tyler Suard feel free to reach out to me directly at: myFirstName.myLastName@microsoft.com. Would be happy to hear any feedback you might have.
Tyler Suard 155 Reputation points

2024-04-05T18:17:11.4866667+00:00

Thank you James, I appreciate your help.
Tyler Suard 155 Reputation points

2024-04-05T18:44:59.64+00:00

James, I upgraded to an M50 tier. Now I am running the index creation from a python script and get an error.

Here is my code, using pymongo:

vector_db.command( { "createIndexes": "WCS_parts_vector_collection", "indexes": [ { "name": "WCS_parts_vector_index", "key": {"Vector": "cosmosSearch"}, "cosmosSearchOptions": { "kind": "vector-hnsw", "efConstruction": 64, # default value "similarity": "COS", "dimensions": 1536, }, } ], } )

After running it, I get this error:

raise ExecutionTimeout(errmsg, code, response, max_wire_version)

pymongo.errors.ExecutionTimeout: The command being executed was terminated due to a command timeout. This may be due to concurrent transactions. Consider increasing the maxTimeMS on the command., full error: {'ok': 0.0, 'errmsg': 'The command being executed was terminated due to a command timeout. This may be due to concurrent transactions. Consider increasing the maxTimeMS on the command.', 'code': 50, 'codeName': 'ExceededTimeLimit'}
James Codella 10 Reputation points Microsoft Employee

2024-04-05T18:57:35.2933333+00:00

Seems like it's just timing out since it's taking a while to build the index on a large number of vectors. You can increase the Collection operation timeout by collection.maxTimeMS(n)

Try setting it to something large (it's in milliseconds)
Tyler Suard 155 Reputation points

2024-04-05T19:07:18.8033333+00:00

Thank you for your help again James, but I am getting this error when I try to do what you recommended:
Tyler Suard 155 Reputation points

2024-04-05T19:12:13.2233333+00:00

Ok I think I figured it out,

vector_db.command(create index params in here).maxTimeMS(1000000000000000)
Tyler Suard 155 Reputation points

2024-04-05T19:12:51.2233333+00:00

No, actually even after that I still got the timeout error. What am I doing wrong?
Tyler Suard 155 Reputation points

2024-04-05T19:21:41.33+00:00

I can't figure out where to execute the maxTimeMS command, even though I am looking everywhere online.
James Codella 10 Reputation points Microsoft Employee

2024-04-05T19:24:48.8+00:00

vector_db.command( { "createIndexes": "WCS_parts_vector_collection", "indexes": [ { "name": "WCS_parts_vector_index", "key": {"Vector": "cosmosSearch"}, "cosmosSearchOptions": { "kind": "vector-hnsw", "efConstruction": 64, # default value "similarity": "COS", "dimensions": 1536, }, } ], }, maxTimeMS=100000000 )

Hmm, sorry for that. Try this, I just verified it should work.

We can update the documentation to provide clearer guidance on how to address the timeout.

Answer 1

In general if your resource runs out of memory, you can try either scaling up the resource to a higher cluster tier, or scaling down the index parameters (e.g., for HNSW you can reduce m or efConstruction and this will reduce the memory footprint of the index. However, you may lose some accuracy in doing so.)

For more than ~10k 1536-dimensional vectors, I wouldn't recommend anything lower than M40. Also, if you scale up to an M40 or above you'll be able to create an HNSW index. HNSW should offer you better latency, throughput, and accuracy compared to IVF at the scale you have.

For example, I was able to have >1M vectors on an M50 without issue using HNSW.

Share via

Error when creating Vector Index for MongoDB vCore

1 answer

Your answer