KLIP 7 - Kafka Connect Integration
Author: agavra | Release Target: 5.4 | Status: Approved | Discussion: link
tl;dr: provide first-class integration with Kafka connect for KSQL data ingress and egress
Motivation and background
Databases provide three functions:
- Data ingress (i.e. inserting data into the database)
- Data transform (i.e. deriving new data from old)
- Data egress (i.e. querying and/or ETL)
In order to develop KSQL as the streaming database, we must provide a complete story around each of these functionality. Today, KSQL is a compelling tool to accomplish data transformations but relies on external mechanisms to import and export data. This KLIP lays out foundations for an integration with Kafka Connect, a system that provides powerful primitives for data ingress and egress for Apache Kafka.
This document will extend the canonical KSQL example of joining pageviews with a user table, but in this document we will assume that the user table is stored in a JDBC compatible database, as opposed to a Kafka changelog topic. After implementing this KLIP, it should be possible for a KSQL user to perform a stream-table join on these two without leaving KSQL's console.
This KLIP will:
- define a SQL-like language for basic connector management
- provide a guiding principles and high level architecture for a connect-ksql integration
This KLIP will not:
- propose improvements to Kafka Connect such as easing the configuration burden, but will discuss how to leverage future Connect improvements
This proposal targets usability improvements for KSQL and simplifies bootstrapping KSQL for users that may already have some, but not all, of their data in Kafka. As with the motivating example, KSQL users will be able to easily model external systems within KSQL. The target user has already played around with KSQL and is working on prototyping their first real application with existing data.
There are three ways to integrate KSQL with Connect:
- Run them separately and supply syntax to communicate between them
- Embed Connect into KSQL, managing individual connect workers as part of the KSQL ecosystem
- Package Connect and KSQL into a single JVM, but just spin them up at the top level
There is some consensus that at scale KSQL and Connect should run as separate components. This means that the API that we design will not be able to leverage any "embedded-ness" of Connect, nullifying most of the benefits of the second approach.
There are still, however, benefits to packaging connect and KSQL in the same JVM for a seamless experience for new users. Namely, new users will only need to download a single package, spin up a single process and maintain only a single application. There are some (solvable) concerns with this approach, but for the reason outlined above, the syntax will be developed with the assumption that Connect is running as a separate component.
The recommendation is to support 1 & 3, tabling 2 unless a need for it comes in the future
Analogous to DDL and DML, the Connect syntax will separate actions that have side effects with actions that simply declare metadata on top of existing data. This solves two issues:
- A single connector can have many topics, which makes modeling them as a single statement difficult
- A connector can be modeled in different ways (e.g. stream/table) and connector lifecycle should be independent with how KSQL interprets it
There are a few more minor points of consideration:
- All "DML" commands that cause changes in connect will not be submitted to the command topic; instead they will be issued on the REST server (this follows the pattern that all external-state modifications are not distributed)
- Some connectors will have KSQL-specific wrapper implementations that will allow simpler configuration and tighter integration. These connectors will also automatically issue the DDL statements for any topics that they create, when possible.
The first step is to create and start a Connector via the REST API (
ksql> CREATE SOURCE CONNECTOR my-postgres-jdbc WITH ( \ connector.class="io.confluent.connect.jdbc.jdbcSourceConnector", \ tasks.max=1, \ connection.url=jdbc:postgresql://localhost:5432/my-db, \ ...); -------------- SOURCE CREATED --------------
Note: as Connect manages to simplify the configuration, whether through templates or any other means, we will immediately be able to leverage this simplification in KSQL's WITH clause.
Next, we provide tools to understand the status of the connector and corresponding topics hitting
GET /connectors and
GET /connectors/<name> REST endpoints:
ksql> SHOW CONNECTORS; name | connector | topic prefix -----------------|-----------------------------------------------|-------------------- my-postgres-jdbc | io.confluent.connect.jdbc.JdbcSourceConnector | test-postgres-jdbc ksql> DESCRIBE my-postgres-jdbc; name: my-postgres-jdbc connector: io.confluent.connect.jdbc.SourceConnector topic-prefix: test-postgres-jdbc worder id | task | status -----------------|------|--------- 10.200.7.69:8083 | 0 | RUNNING name | topic | schema -----------|-------------------------------|-------------------------------------- user-table | test-postgres-jdbc-user-table | id VARCHAR, user VARCHAR, age INTEGER page-table | test-postgres-jdbc-page-table | id VARCHAR, page VARCHAR, url VARCHAR
Then the user can create streams or tables from the underlying topics (this can happen automatically for some set of blessed connectors):
ksql> CREATE TABLE users_original WITH (kafka_topic = 'test-postgres-jdbc-user-table', ...); -------------- STREAM CREATED --------------
At this point, assuming the
pageviews_original stream is already created, we can join the newly
users_original table with the
pageviews_original stream. Finally, the user can drop and
terminate the connectors.
ksql> DROP CONNECTOR my-postgres-jdbc;
Data egress would work in a similar fashion:
ksql> CREATE SINK CONNECTOR my-elastic-sink WITH (sources='my-table1', ...); ksql> SHOW SINK CONNECTORS; name | connector | topic prefix -----------------|-----------------------------------------------|-------------------- my-elastic-sink | io.confluent.connect.elasticsearch.ElasticsearchSinkConnector | elastic-sink ksql> DESCRIBE my-elastic-sink; name: my-elastic-sink connector: io.confluent.connect.elasticsearch.ElasticsearchSinkConnector topic-prefix: elstic-sink worder id | task | status -----------------|------|--------- 10.200.7.69:8083 | 1 | RUNNING name | topic -------------------------------------------- elasticsearch1 | elastic-sink-elasticsearch1 ksql> DROP CONNECTOR my-elastic-sink TERMINATE;
This section is a record of some suggestions for Connect that would make it more KSQL-friendly:
- Exposing Topics & Schemas (P0): as it stands, Connect does not expose the topics or schemas
for the connectors it manages. This means that, at best, we need to hack together logic that
extracts information from a connect configuration and scan Kafka for matching topics (e.g. for
JDBC connect we would need to scan all topics that start with
topic.prefix) and at worst it is impossible (the topic names are not necessarily determinable from the config). Similarly with schemas, at best we can scan Schema Registry for the schema of the topic, but at worst there is no guarantee that the topic uses AVRO or is registered in schema registry using the standard subject naming schema (e.g.
<topic>-value), rendering a schema unrecoverable programmatically. For true KSQL-Connect integration, we would need an exposed
/connectors/<name>/topicsendpoint and a
/connectors/<name>/<topic>/schemaendpoint (or something similar).
- JDBC Primary Key (P1): Since we're targeting the JDBC connector as the first "blessed" integration point with KSQL, it is important that the experience is as good as possible. It would be great if the connector populated the primary key of the row into the kafka record without requiring an SMT.
- Push updates (P2): It would be nice to be able to "subscribe" to get updates for any new topic/schema that was created (see the first item) without polling connect.
- Error Handling: it is difficult to understand whether a given connector is or is not working, and further difficult to debug why it is not working. This is out of scope for this KLIP and should be addressed in connect and then exposed externally. There is a limited set of synchronous errors that can be easily identified (e.g. missing configurations) that can be bubbled up directly to the user at the time when they submit the request.
- Secret Management: some connectors (e.g. JDBC connector) may require passing secrets to
access the database. For the MVP, the security will naively rely on HTTPS and Connect - requiring
users to specify their secrets in the SQL statement (e.g.
CREATE ... WITH(foo.username='me', foo.password='secret')). Although this is not ideal, these command will not be distributed onto the KSQL command topic. It is recommended that the Connect configuration topic, though, is locked down to prevent these secrets from leaking there. In a later iteration, KSQL can integrate more directly with Connect security features by passing tokens to some secret manager (e.g. Vault) and resolving them only server-side.
The test plan will involve writing a full demo application for the motivating use case, where the user table is stored in Postgres. Beyond that, integration tests that spin up connect will be added to ensure that compatibility does not degrade across releases.
See section on "Secret Management"