pg2kinesis uses logical decoding in Postgres 9.4 or later to capture a consistent, continuous stream of events from the database and publishes them to an AWS Kinesis stream in a format of your choosing.
It does this without requiring any changes to your schema like triggers or "shadow" columns or tables, and has a negligible impact on database performance. This is done while being extremely fault tolerant. No data loss will be incurred on any type of underlying system failure including process crashes, network outages, or ec2 instance failures. However, in these situations there will likely be records that are sent more than once, so your consumer should be designed with this in mind.
The fault tolerance comes from guarantees provided by the underlying technologies and from the "2-phase commit" style of publishing inherent to the design of the program. Changes are first peeked from the replication slot and published to Kinesis. Once Kinesis successfully recieves a batch of records, we advance the xmin of the slot, thereby telling postgres it is safe to reclaim the space taken by the wal logs. As is always the case with logical replication slots, unacknowledged data on the slot will consume disk on the database until it is read.
There are other utilities that do similar things, often by injecting a C library into Postgres to do data transformations in place. Unfortunately these approaches are not suitable for managed databases like AWS' RDS where support for various plugins is limited and ultimately determined by the hosting provider. We specifically created pg2kinesis to make use of logical decoding on Amazon's RDS for PostgreSQL. Amazon only supports logical decoding if you use the built in test_decoding output plugin. What is special about our utility is it takes the output of the test_decoding plugin, transforms it based on a formatter you can define, guarantees publishing to a Kinesis stream in transaction commit time order and with a guarantees that no data will be lost.
Prerequisites:
- Python 2.7.*
- AWS-CLI installed and configured
- A PostgreSQL 9.4+ server with logical replication enabled
- A Kinesis stream
Install:
pip install pg2kinesis
To run tests you will need a clone of the repo and have to install some additional requirements:
git clone [email protected]:handshake/pg2kinesis.gitcd pg2kinesispip install -r requirements.txtpytest
Run pg2kinesis --help to get a list of the latest command line options.
By default pg2kinesis attempts to connect to a local postgres instance and publish to a stream named pg2kinesis using the AWS credentials in the environment the utility was invoked in.
On successful start it will query your database for the primary key definitions of every table in --pg-dbname. This is used to identify the correct column in the test_decoding output to publish. If a table does not have a primary key its changes will NOT be published.
You have the choice for 2 different textual formats that will be sent to the kinesis stream:
CSVFormatter: outputs stings to Kinesis that look like:0,CDC,<transaction_id (xid)>,<table name>,<dml operation:DELETE|INSERT|UPDATE>,<primary key of row>
CSVPayloadFormatter: outputs similar to the above except the 3rd column is now a json object representing the change.:0,CDC,{ "xid": <transaction_id> "table": <...> "operation: <...> "pkey": <...> }
pg2kinesis is based on the ideas of others including:
- Logical Decoding: a new world of data exchange applications for Postgres SQL [(slides)]
- psycopg2 [(main]) [(repo)]
- bottledwater-pg [(blog)] [(repo)]
- wal2json [(repo)]
- Allow payload format to be specified on Command Line.
- Allow HUPing to notify utility to regenerate primary key cache
- Support above on a schedule specified via commandline with sensible default of once an hour.
- Full Row information publishing (currently we only emit the DML, primary key and table on any type of change)
- Python 3 Support