A CDC pipeline that streams postgres database table changes to BigQuery via Debezium, PubSub, Avro, Dataflow+Python...
At least, that's the original plan. Now this repo is just a mix bag of experiments with Apache Beam.
# Python 3.8 is not supported in Beam 2.21 and Dataflow. To install Python 3.7 in Ubuntu 20.04 you can do:
sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt-get update
sudo apt-get install python3.7
virtualenv --python=python3.7 ~/py37
source ~/py37/bin/activate
# jq is not required, but nice to have
sudo apt install docker.io docker-compose jq
# Start debezium + example postgres db
docker-compose up
# Remove containers. Append --volumes to drop volumes as well
docker-compose down
# Setup/update connector
curl -i -X DELETE http://localhost:8083/connectors/inventory-connector \
&& curl -i -X POST -H "Accept:application/json" \
-H "Content-Type:application/json" http://localhost:8083/connectors/ \
-d @register-postgres.json
# Query available connector
curl -H "Accept:application/json" localhost:8083/connectors/
# See inventory customer schema in connector
curl -sH "Accept:application/json" localhost:8083/connectors/inventory-connector | jq
# See customer schema in schema registry
curl -sX GET http://localhost:8081/subjects/dbserver1.inventory.customers-value/versions/1 | jq '.schema | fromjson'
# See schema from registry from global id
curl -sX GET http://localhost:8081/schemas/ids/1 | jq '.schema | fromjson'
# Access postgres database
psql postgresql://postgres:postgres@localhost:5432/postgres
# ...you can also access from within the docker container
docker-compose exec postgres bash -c 'psql -U postgres postgres'
# Start test kafka client
pip install -r requirements.txt
python kafka-client.py
# -----------------------------------------------------------------------------------
# Set env for pubsub to run locally
export PUBSUB_EMULATOR_HOST=localhost:8085
# Start pubsub client
python pubsub-client.py
# Kotlin (uses Kafka instead of pubsub) ---------------------------------------------
# Direct Runner (--output is optional and writes to files)
mvn compile exec:java \
-Dexec.mainClass=bruceoutdoors.beam.examples.PostgresCDCBigQuery \
-Dexec.args="--output=written-records \
--project=crafty-apex-264713
"
# Dataflow Runner (Must be executed from where it can access Kafka)
# Doesn't work... java.io.IOException: Reader-0: Timeout while initializing partition
# 'dbserver1.inventory.customers-0'. Kafka client may not be able to connect to servers.
mvn compile exec:java \
-P dataflow-runner \
-Dexec.mainClass=bruceoutdoors.beam.examples.PostgresCDCBigQuery \
-Dexec.args="--runner=DataflowRunner \
--project=crafty-apex-264713 \
--region=asia-east1 \
--tempLocation=gs://kakfa-testing-bucket/tmp \
--stagingLocation=gs://kakfa-testing-bucket/staging \
--schemaRegistry=http://10.140.0.4:8081 \
--bootstrapServers=http://10.140.0.4:9092
"
# Flink Runner (Experimental!)
mvn compile exec:java \
-P flink-runner \
-Dexec.mainClass=bruceoutdoors.beam.examples.PostgresCDCBigQuery \
-Dexec.args="--runner=FlinkRunner \
--flinkMaster=localhost:8082 \
--project=crafty-apex-264713 \
--schemaRegistry=http://localhost:8081 \
--bootstrapServers=http://localhost:9092
"
# Python -----------------------------------------------------------------------------
# Direct Runner (You may want to comment out BigQuery task)
python postgres-bigquery-beam.py \
--failed-bq-inserts failed-inserts \
--project crafty-apex-264713
# Run in job in Dataflow:
export GOOGLE_APPLICATION_CREDENTIALS=/home/bruce/secret_gcp.json
python postgres-bigquery-beam.py \
--runner DataflowRunner \
--project crafty-apex-264713 \
--region asia-east1 \
--temp_location gs://kakfa-testing-bucket/tmp \
--staging_location gs://kakfa-testing-bucket/staging \
--failed-bq-inserts gs://kakfa-testing-bucket/failed_inserts \
--schema_registry 'http://10.140.0.4:8081' \
--requirements_file dataflow-requirements.txt
I've initially used confluent-kafka[avro], but because it requires some non-PyPi setup which I simply could not get to work. I've since switched to use python-schema-registry-client. Figuring this out is notoriously hard since you had to deliberately go to the Job worker logs in a separate pop-up to see the error messages - in the job monitor itself the error is nowhere to be found, but the job seems to be running without producing any results.