Kafka to BigQuery Sink
- Consumer: Consumes messages from kafka in batches, and pushes these batches to Read & Commit queues. These queues are blocking queues, i.e, no more messages will be consumed if the queue is full. (This is configurable based on poll timeout)
- BigQuery Worker: Polls messages from the read queue, and pushes them to BigQuery. If the push operation was successful, BQ worker sends an acknowledgement to the Committer.
- Committer: Committer receives the acknowledgements of successful push to BigQuery from BQ Workers. All these acknowledgements are stored in a set within the committer. Committer polls the commit queue for message batches. If that batch is present in the set, i.e., the batch has been successfully pushed to BQ, then it commits the max offset for that batch, back to Kafka, and pops it from the commit queue & set.
- A kafka cluster which has messages pushed in proto format, which beast can consume
- should have BigQuery project which has streaming permission
- create a table for the message proto
- create configuration with column mapping for the above table and configure in env file
- env file should be updated with bigquery, kafka, and application parameters
git clone https://github.com/gojekfarm/beast
export $(cat ./env/sample.properties | xargs -L1) && gradle clean runConsumer
The image is available in gojektech dockerhub.
export TAG=80076c77dc8504e7c758865602aca1b05259e5d3
docker run --env-file beast.env -v ./local_dir/project-secret.json:/var/bq-secret.json -it gojektech/beast:$TAG
-v
mounts local secret fileproject-sercret.json
to the docker mentioned location, andGOOGLE_CREDENTIALS
should match the same/var/bq-secret.json
which is used for BQ authentication.TAG
You could update the tag if you want the latest image, the mentioned tag is tested well.
Create a beast deployment for a topic in kafka, which needs to be pushed to BigQuery.
- Deploymet can have multiple instance of beast
- A beast container consists of the following threads:
- A kafka consumer
- Multiple BQ workers
- A committer
- Deployment also includes telegraf container which pushes stats metrics Follow the instructions in chart for helm deployment
Given a TestMessage proto file, you can create bigtable with schema
# create new table from schema
bq mk --table <project_name>:dataset_name.test_messages ./docs/test_messages.schema.json
# query total records
bq query --nouse_legacy_sql 'SELECT count(*) FROM `<project_name>:dataset_name.test_messages LIMIT 10'
# update bq schema from local schema json file
bq update --format=prettyjson <project_name>:dataset_name.test_messages booking.schema
# dump the schema of table to file
bq show --schema --format=prettyjson <project_name>:dataset_name.test_messages > test_messages.schema.json
You can generate messages with TestMessage.proto with sample-kafka-producer, which pushes N messages
- You could raise issues or clarify the questions
- You could raise a PR for any feature/issues To run and test locally:
git clone https://github.com/gojekfarm/beast
export $(cat ./env/sample.properties | xargs -L1) && gradlew test
- You could help us with documentation