New in CockroachDB 24.2 is support for vectors. Now we can store vector
embeddings within CockroachDB with pgvector-compatible semantics to build
AI-driven applications. Numerous built-in functions have been added for running
similarity search across vectors. Here, we use the <=> (cosine similarity)
operator.
The caveat with this is that vector indexing is not supported in this release. Still, that won't stop us from experimenting with some semantic indexing and search. To make up for the lack of indexing on these vectors, we'll add a K-Means clustering step to our app and create an index on the cluster ID values which we'll map to the primary key of the table storing the data. This gives us a two phased approach to search: (1) use this index, (2) order the resulting rows according to each row's cosine similarity to the query string.
This demo can be run locally using the steps outlined below or in Kubernetes (K8s), adjacent to a CockroachDB deployment, using the deployment defined here.
- Install a CockroachDB instance, using version 24.2+
- Apply a CockroachDB Enterprise license
- Create a user account for the app
- GRANT this user/role access to the DB being used
- The app creates all the required tables and indexes
- Run the following to prevent failures from uploading K-Means models to the DB's
blob_storetable:SET CLUSTER SETTING sql.conn.max_read_buffer_message_size = '64 MiB';
The simplest way to get started is to run the Docker image. To do that, edit the
./docker_run_image.sh file, adjusting the lines near the top as necessary:
# EDIT THESE TO SUIT YOUR DEPLOYMENT
crdb_certs="$HOME/certs"
export DB_URL="postgres://test_role:123abc@host.docker.internal:26257/defaultdb?sslmode=require&sslrootcert=$crdb_certs/ca.crt"
port=1999 # You point your client at this port on localhost, so 'export FLASK_PORT=1999' in env.sh
You can then skip down to Start the Flask server Docker image.
If deploying locally or on a VM, edit the ./env.sh file; if deploying in K8s, these
variables are defined in the deployment YAML file (./k8s/crdb-embeddings.yaml).
Here is a detailed explanation of the various environment variables used by the app
and related scripts (search, index, ...):
Host name and port used by the app (a Python Flask REST service):
export FLASK_PORT=8080
export FLASK_HOST=some-host.domain
The DB connection string:
export DB_URL="postgres://some_role:password@127.0.0.1:26257/db_name?sslmode=require&sslrootcert=/certs/ca.crt"
Log level (INFO is fairly chatty):
export LOG_LEVEL=INFO
Number of threads to run in the Flask app:
export N_THREADS=10
The minimum length of a sentence, in characters:
export MIN_SENTENCE_LEN=8
The number of clusters to create in the K-Means model. The tradeoff here is query speed vs. recall; e.g. if the number is too low, more data must be scanned during the cosine similarity phase. If the number is too high, then matching documents may be missed entirely as their cluster ID value will not align with the cluster ID value that gets mapped to the query string. For the small data size used in these experiments, 50k rows, a value of 500 seemed like the best fit. For a larger data set, experiments showed using a value of about 1.3% of the number of rows was a good tradeoff:
export N_CLUSTERS=500
The fraction of rows to scan when building the K-Means model. A value of 1.0 might make more sense for a small data set, but a smaller fraction would be better suited to a larger data set. Again, there's a tradeoff here in terms of fidelity, with a larger fraction (theoretically) leading to a better model. Again, when the data set became larger, experiments showed that a fraction corresponding to about 35k rows in the table was a sweet spot:
export TRAIN_FRACTION=0.75
A filesystem location where the model can be read and written. This is used during the "bootstrap" phase when the model file is downloaded from the URL provided below. After that, the model is stored in the DB itself:
export MODEL_FILE=/tmp/model.pkl
This model was built according to the discussion above, and it should be suitable for getting started:
export MODEL_FILE_URL="https://storage.googleapis.com/crl-goddard-text/model_Fastembed_500.pkl"
There are a few more uploaded, built from increasingly larger data sets, where the numeric
value before the .pkl suffix is the N_CLUSTERS value used:
This applies during the process of assigning a cluster ID value to each of the rows. The value of 1024 shown here yielded the best data insert rate:
export BATCH_SIZE=1024
When building the K-Means model, the verbosity level can be adjusted. The value of "1" here yields enough output so you can be convinced something is happening. It can be disabled by setting this to "0":
export KMEANS_VERBOSE=1
Building a K-Means model is iterative. This value of "100" seems to work well enough:
export KMEANS_MAX_ITER=100
This is just a string that's required as a URL parameter to the app's /cluster_assign
and /build_model endpoints since, if the app is exposed on a network, we'd rather
not have random clients hitting these. Set it to any string containing no spaces or special
charaters:
export SECRET="TextWithNoSpecialChars"
The blob_store table gets a new row containing a model each time that build_model
process runs. This table is pruned to ensure it has at most this number of rows:
export BLOB_STORE_KEEP_N_ROWS=3
This is rumored to prevent a warning from the HuggingFace library about forking multiple threads, but it ended up having no impact on that. Still, I'll keep it for the sake of documentation:
export TOKENIZERS_PARALLELISM=false
The Fastembed TextEmbedding model relies on some C++ libraries which attempt to grab an unbounded amount of memory and, so far, I haven't taken the time to find a way to prevent this. This setting keeps Python to a bounded amount of memory for data which will result in periodic memory allocation failures of the form
2024-09-05 19:56:10.642871010 [E:onnxruntime:, sequential_executor.cc:516 ExecuteKernel] Non-zero status code returned while running Attention node. Name:'Attention_0' Status Message: /onnxruntime_src/onnxruntime/core/framework/bfc_arena.cc:376 void* onnxruntime::BFCArena::AllocateRawInternal(size_t, bool, onnxruntime::Stream*, bool, onnxruntime::WaitNotificationFn) Failed to allocate memory for requested buffer of size 3221225472
This is the environment variable:
export MEMORY_LIMIT_MB=6144
Some texts on Wikipedia were immense, so this value is used to manage the number of sentences that will be indexed (each sentence becomes a chunk):
export MAX_CHUNKS=256
The number of results retrieved and displayed as JSON by the client:
export MAX_RESULTS=5
When starting out without a K-Means model (if not using the MODEL_FILE_URL, for
example), this would be set to True and then the documents could be added using
index_doc.py and then the model would be built as shown below. Then, the app
would be restarted with this value reset to False and the cluster assignment step
would need to be done prior to running a search.
export SKIP_KMEANS=False
[13:22:56 crdb-embeddings]$ ./docker_run_image.sh
1.26: Pulling from mgoddard/crdb-embeddings-arm
Digest: sha256:a83a9d8cd33342eba584ffa08a72660ad4595171ee8e4dfa78bf9a899de2315e
Status: Image is up to date for mgoddard/crdb-embeddings-arm:1.26
docker.io/mgoddard/crdb-embeddings-arm:1.26
Collecting psycopg2-binary
Downloading psycopg2_binary-2.9.9-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl (2.9 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 2.9/2.9 MB 16.6 MB/s eta 0:00:00
[... Many more such lines removed here for brevity ...]
Building wheels for collected packages: PyStemmer
Building wheel for PyStemmer (setup.py): started
Building wheel for PyStemmer (setup.py): finished with status 'done'
Created wheel for PyStemmer: filename=PyStemmer-2.2.0.1-cp38-cp38-linux_aarch64.whl size=612716 sha256=9b9eb0b5d11052fff130237e588c074cef726edcd7dcc246301a43d28a6478fe
Stored in directory: /tmp/pip-ephem-wheel-cache-aj__ssz4/wheels/78/04/32/a81f10f01775fcadba622dbcf8305f8053ab1db21b20a25fc4
Successfully built PyStemmer
Installing collected packages: snowballstemmer, PyStemmer, mpmath, mmh3, flatbuffers, zipp, waitress, urllib3, typing-extensions, tqdm, threadpoolctl, sympy, pyyaml, psycopg2-binary, protobuf, pillow, packaging, numpy, MarkupSafe, loguru, joblib, itsdangerous, idna, humanfriendly, greenlet, fsspec, filelock, click, charset-normalizer, certifi, Werkzeug, SQLAlchemy, scipy, requests, psycopg2-pool, pgvector, onnx, Jinja2, importlib-metadata, coloredlogs, sqlalchemy-cockroachdb, scikit-learn, onnxruntime, huggingface-hub, Flask, tokenizers, fastembed
Successfully installed Flask-2.1.0 Jinja2-3.1.4 MarkupSafe-2.1.5 PyStemmer-2.2.0.1 SQLAlchemy-2.0.25 Werkzeug-2.2.2 certifi-2024.7.4 charset-normalizer-3.3.2 click-8.1.7 coloredlogs-15.0.1 fastembed-0.3.6 filelock-3.15.4 flatbuffers-24.3.25 fsspec-2024.6.1 greenlet-3.0.3 huggingface-hub-0.24.6 humanfriendly-10.0 idna-3.8 importlib-metadata-8.4.0 itsdangerous-2.2.0 joblib-1.4.2 loguru-0.7.2 mmh3-4.1.0 mpmath-1.3.0 numpy-1.24.4 onnx-1.16.2 onnxruntime-1.19.0 packaging-24.1 pgvector-0.3.2 pillow-10.4.0 protobuf-5.27.4 psycopg2-binary-2.9.9 psycopg2-pool-1.2 pyyaml-6.0.2 requests-2.32.3 scikit-learn-1.3.2 scipy-1.10.1 snowballstemmer-2.2.0 sqlalchemy-cockroachdb-2.0.1 sympy-1.13.2 threadpoolctl-3.5.0 tokenizers-0.20.0 tqdm-4.66.5 typing-extensions-4.12.2 urllib3-2.2.2 waitress-3.0.0 zipp-3.20.1
kmeans_max_iter: 250 (set via 'export KMEANS_MAX_ITER=25')
kmeans_verbose: 1 (set via 'export KMEANS_VERBOSE=1')
skip_kmeans: False (set via 'export SKIP_KMEANS=False')
batch_size: 768 (set via 'export BATCH_SIZE=512')
n_clusters : 1000 (set via 'export N_CLUSTERS=50')
train_fraction: 0.5 (set via 'export TRAIN_FRACTION=0.10')
model_file: /tmp/model.pkl (set via 'export MODEL_FILE=./model.pkl')
model_url: https://storage.googleapis.com/crl-goddard-text/model_Fastembed_1k.pkl (set via 'export MODEL_FILE_URL=https://somewhere.com/path/model.pkl')
min_sentence_len: 8 (set via 'export MIN_SENTENCE_LEN=12')
n_threads: 10 (set via 'export N_THREADS=10')
max_retries: 3 (set via 'export MAX_RETRIES=3')
shared secret: TextWithNoSpecialChars
blob_store_keep_n_rows: 3
Log level: INFO (export LOG_LEVEL=[DEBUG|INFO|WARN|ERROR] to change this)
Fetching 5 files: 100%|██████████| 5/5 [00:03<00:00, 1.48it/s]
[08/28/2024 05:27:56 PM MainThread] TextEmbedding model ready: 4.718861103057861 s
[08/28/2024 05:27:56 PM MainThread] Checking whether text_embed table exists
[08/28/2024 05:27:57 PM MainThread] text_embed table already exists
/usr/local/lib/python3.8/site-packages/sqlalchemy_cockroachdb/base.py:226: SAWarning: Did not recognize type 'vector' of column 'embedding'
warn(f"Did not recognize type '{type_name}' of column '{name}'")
[08/28/2024 05:27:57 PM MainThread] Fetching model from the DB ...
/usr/local/lib/python3.8/site-packages/sklearn/base.py:348: InconsistentVersionWarning: Trying to unpickle estimator KMeans from version 1.5.1 when using version 1.3.2. This might lead to breaking code or invalid results. Use at your own risk. For more info please refer to:
https://scikit-learn.org/stable/model_persistence.html#security-maintainability-limitations
warnings.warn(
[08/28/2024 05:27:58 PM MainThread] OK
[08/28/2024 05:27:58 PM MainThread] K-means model loaded
[08/28/2024 05:27:58 PM MainThread] You may need to update K-means cluster assignments by making a GET request to the /cluster_assign/TextWithNoSpecialChars endpoint.
[08/28/2024 05:27:58 PM MainThread] Serving on http://0.0.0.0:18080
huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
To disable this warning, you can either:
- Avoid using `tokenizers` before the fork if possible
- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
Note: This process is time consuming and would benefit from being parallelized.
The app itself is multi-threaded, and CockroachDB scales very nicely, so altering
./index_doc.py so it can submit multiple indexing requests in parallel is all that
has to be done.
$ . ./env.sh
$ time ./index_doc.py ./data/*.txt
data/5am_club.txt: SUCCESS (t = 1010.720 ms)
data/aus_nz.txt: SUCCESS (t = 323.854 ms)
data/bafta_awards.txt: SUCCESS (t = 179.528 ms)
data/bourdain.txt: SUCCESS (t = 322.123 ms)
[...]
data/top_jazz_albums.txt: SUCCESS (t = 429.001 ms)
data/undoctored.txt: SUCCESS (t = 491.508 ms)
data/victoria.txt: SUCCESS (t = 539.466 ms)
real 0m33.947s
user 0m0.135s
sys 0m0.048s
[16:35:11 crdb-embeddings]$ q="which amplifiers have vacuum tube triode sound"
mgoddard@crdb-embeddings:~/crdb-embeddings$ time ./search_client.sh $q
[
{
"uri": "data/decware_amp.txt",
"score": 0.7652439003492352,
"chunk": "There are 3 models of our 2 watt series of Zen Triode amplifier."
},
{
"uri": "https://en.wikipedia.org/wiki/Fender_Pro_Reverb",
"score": 0.6877914108857055,
"chunk": "The Pro Reverb is a 40-watt tube amplifier and has a pair of 12\" speakers; however, models later than 1976 were increased to 70 watts."
},
{
"uri": "https://en.wikipedia.org/wiki/Heathkit",
"score": 0.6847249097712207,
"chunk": "Heathkit stereo preamplifier (AA-141) (1962)."
},
{
"uri": "https://en.wikipedia.org/wiki/Heathkit",
"score": 0.6842641030471444,
"chunk": "Wiring of a Heathkit stereo preamplifier (AA-141) (1962)."
},
{
"uri": "https://en.wikipedia.org/wiki/Fender_Pro_Reverb",
"score": 0.6832191377673992,
"chunk": "Teagle and Sprung, authors of the definitive book on Fender amplifiers (Fender Amps: The First Fifty Years) described the Blackface Pro Reverb as the best amplifier ever produced likely due to its combination of the two twelve inch speakers, its high quality reverb and tremolo, and just the right amount of power to get either classic Blackface Fender cleans or natural tube breakup at reasonable volumes."
}
]
real 0m0.101s
user 0m0.008s
sys 0m0.016s
Caveat: If your data set is small, the default model will perform better. For now, if you perform this step and want to roll back to the default model, perform the actions listed under "Rollback to bootstrap model", below.
This takes a while (the TRAIN_FRACTION value affects the time).
[16:55:59 crdb-embeddings]$ . ./env.sh
[16:56:02 crdb-embeddings]$ ./build_model.sh $SECRET
This also takes a while, but is necessary after the model is rebuilt. If the search results don't make sense, it's likely you need to run this.
[16:57:43 crdb-embeddings]$ . ./env.sh
[16:57:50 crdb-embeddings]$ ./cluster_assign.sh $SECRET
In the database, truncate the blob_store table:
defaultdb=> TRUNCATE TABLE blob_store;
Remove the existing model on the filesystem:
[18:49:47 crdb-embeddings]$ . ./env.sh
[18:49:53 crdb-embeddings]$ rm -f $MODEL_FILE
Restart the app
- https://www.cockroachlabs.com/docs/releases/v24.2#v24-2-0-sql
- https://stackoverflow.blog/2023/11/09/an-intuitive-introduction-to-text-embeddings/
- https://huggingface.co/blog/bert-101
- https://huggingface.co/distilbert/distilbert-base-uncased
- https://mccormickml.com/2019/05/14/BERT-word-embeddings-tutorial/
- https://github.com/qdrant/fastembed
- Models supported by Fastembed
- https://github.com/pgvector/pgvector-python?tab=readme-ov-file#psycopg-2
- Example CockroachDB query plan for these semantic searches
- JSON Wikipedia dump from 2020
- https://docs.sqlalchemy.org/en/20/core/connections.html#streaming-with-a-dynamically-growing-buffer-using-stream-results