notai-tech / fastdeploy

Deploy DL/ ML inference pipelines with minimal extra code.

fastDeploy

easy and performant micro-services for Python Deep Learning inference pipelines

• Deploy any python inference pipeline with minimal extra code
• Auto batching of concurrent inputs is enabled out of the box
• no changes to inference code (unlike tf-serving etc), entire pipeline is run as is
• Promethues metrics (open metrics) are exposed for monitoring
• Auto generates clean dockerfiles and kubernetes health check, scaling friendly APIs
• sequentially chained inference pipelines are supported out of the box
• can be queried from any language via easy to use rest apis
• easy to understand (simple consumer producer arch) and simple code base

Installation:

pip install --upgrade fastdeploy fdclient
# fdclient is optional, only needed if you want to use python client

CLI explained

Start fastDeploy server on a recipe:

# Invoke fastdeploy 
python -m fastdeploy --help
# or
fastdeploy --help

# Start prediction "loop" for recipe "echo"
fastdeploy --loop --recipe recipes/echo

# Start rest apis for recipe "echo"
fastdeploy --rest --recipe recipes/echo

Send a request and get predictions:

• Python client usage

• curl usage

• Nodejs client usage

auto generate dockerfile and build docker image:

# Write the dockerfile for recipe "echo"
# and builds the docker image if docker is installed
# base defaults to python:3.8-slim
fastdeploy --build --recipe recipes/echo

# Run docker image
docker run -it -p8080:8080 fastdeploy_echo

Serving your model (recipe):

• Writing your model/pipeline's recipe

Where to use fastDeploy?

• to deploy any non ultra light weight models i.e: most DL models, >50ms inference time per example
• if the model/pipeline benefits from batch inference, fastDeploy is perfect for your use-case
• if you are going to have individual inputs (example, user's search input which needs to be vectorized or image to be classified)
• in the case of individual inputs, requests coming in at close intervals will be batched together and sent to the model as a batch
• perfect for creating internal micro services separating your model, pre and post processing from business logic
• since prediction loop and inference endpoints are separated and are connected via sqlite backed queue, can be scaled independently

Where not to use fastDeploy?

• non cpu/gpu heavy models that are better of running parallely rather than in batch
• if your predictor calls some external API or uploads to s3 etc in a blocking way
• io heavy non batching use cases (eg: query ES or db for each input)
• for these cases better to directly do from rest api code (instead of consumer producer mechanism) so that high concurrency can be achieved