This repository contains the source code of the Kafka Connect Kusto sink connector.
"Kusto" is the Microsoft internal project code name for Azure Data Explorer, Microsoft Azure's big data analytical database PaaS offering.
1. Overview
2. Integration design
3. Features supported
4. Connect worker properties
5. Sink properties
6. Roadmap
7. Deployment overview
8. Connector download/build from source
9. Test drive the connector - standalone mode
10. Distributed deployment details
11. Test drive the connector - distributed mode
12. Apache Kafka version - Docker Kafka Connect base image version - Confluent Helm chart version related
13. Other
14. Need Support?
15. Major version specifics
16. Release History
17. Contributing
Azure Data Explorer is a first party Microsoft big data analytical database PaaS, purpose built for low latency analytics of all manners of logs, all manners of telemetry and time series data.
Kafka ingestion to Azure Data Explorer leverages Kafka Connect. Kafka Connect is an open source Apache Kafka ETL service for code-free, configuration based, scalable, fault tolerant integration with Kafka from/to any system through development of connector plugins (integration code). The Azure Data Explorer team has developed a sink connector, that sinks from Kafka to Azure Data Explorer.
Our connector is gold certified by Confluent - has gone through comprehensive review and testing for quality, feature completeness, compliance with standards and for performance.
The connector is open source and we welcome community contribution.
Integration mode to Azure Data Explorer is queued or streaming ingestion leveraging the Azure Data Explorer Java SDK. Events are dequeued from Kafka and per the flush* sink properties, batched or streamed, and shipped to Azure Data Explorer as gzipped files. In case of batch ingestions, Azure Data Explorer has a configurable table ingest batching policy, based on which ingestion into tables occurs automatically. In case of streaming ingestion, configure 'streaming' as 'true' in 'kusto.tables.topics.mapping', by default streaming is set as false.
- The connector fails fast if any of the "required" sink properties are unavailable
- The connector checks for access to the cluster and database table and shuts down if inaccessible
- The connector checks for any dependent properties that are selectively required and shuts down if unavailable (e.g. a converter specified that relies on schema registry but missing a schema registry URL)
- The connector supports configurable
- shut down in the event of an error
OR - log the error and process as many events as possible
- The connector supports retries for transient errors with the ability to provide relevant parameters
- and retries with exponential backoff
- The connector supports Avro, JSON, CSV formats
- It supports Parquet and ORC file formats with the ByteArrayConverter specified below
- The connector supports schema registry for avro and json
- The connector does not support schema evolution currently
- The connector supports the following converters:
| # | Converter | Details |
|---|---|---|
| 1 | org.apache.kafka.connect.storage.StringConverter | Use with csv/json |
| 2 | org.apache.kafka.connect.json.JsonConverter | Use with schemaless json |
| 3 | io.confluent.connect.avro.AvroConverter | Use with avro |
| 4 | io.confluent.connect.json.JsonSchemaConverter | Use with json with schema registry |
| 5 | org.apache.kafka.connect.converters.ByteArrayConverter | Use with ORC, Parquet files written as messages to Kafka |
| 6 | io.confluent.connect.protobuf.ProtobufConverter | Use with protobuf format with schema registry |
- The connector does not support transformers. Prefer transformation on the server side in Kafka or ingestion time in Azure Data Explorer with update policies.
- The connector supports multiple topics to multiple tables configuration per Kafka Connect worker
- The connector supports user provided "Dead Letter Queue", a Kafka Connect construct; E.g. If Avro messages are written to a "dead letter queue" topic that is expecting Json, the avro messages are written to a configurable dead letter queue instead of just being dropped. This helps prevent data loss and also data integrity checks for messages that did not make it to the destination. Note that for a secure cluster, in addition to bootstrap server list and topic name, the security mechanism, the security protocol, jaas config have to be provided for the Kafka Connect worker and in the sink properties
- The connector supports user provided miscelleneous "Dead Letter Queue" for transient and non-deserialization errors (those not managed by Kafka Connect); E.g. If network connectivity is lost to Azure Data Explorer, the connector retries and eventually writes the queued up messages to the miscelleneous "Dead Letter Queue". Note that for a secure cluster, in addition to bootstrap server list and topic name, the security mechanism, the security protocol, jaas config have to be provided for the Kafka Connect worker and in the sink properties
- Azure Data Explorer is an append only immutable database. Infrastructure failures and unavoidable external variables that can lead to duplicates cant be remediated via upsert commands as upserts are not supported.
Therefore the connector supports "At least once" delivery guarantees.
- The connector supports overrides at the sink level if overrides are specified at a Kafka Connect worker level. This is a Kafka Connect feature, not specific to the Kusto connector plugin.
- As with all Kafka Connect connectors, parallelism comes through the setting of connector tasks count, a sink property
- Azure Data Explorer supports Azure Active Directory authentication. For the Kusto Kafka connector, we need an Azure Active Directory Service Principal created and "admin" permissions granted to the Azure Data Explorer database.
- Kafka Connect supports all security protocols supported by Kafka, as does our connector
- See below for some security related config that needs to be applied at Kafka Connect worker level as well as in the sink properties
- There are some core configs that need to be set at the Kafka connect worker level. Some of these are security configs and the (consumer) override policy. These for e.g. need to be baked into the Docker image covered further on in this document-
The below covers Confluent Cloud-
Link to end to end sample
ENV CONNECT_CONNECTOR_CLIENT_CONFIG_OVERRIDE_POLICY=All
ENV CONNECT_SASL_MECHANISM=PLAIN
ENV CONNECT_SECURITY_PROTOCOL=SASL_SSL
ENV CONNECT_SSL_ENDPOINT_IDENTIFICATION_ALGORITHM=https
ENV CONNECT_SASL_JAAS_CONFIG="org.apache.kafka.common.security.plain.PlainLoginModule required username=\"<yourConfluentCloudAPIKey>\" password=\"<yourConfluentCloudAPISecret>\";"
The below covers HDInsight Kafka with Enterprise Security Package (Kerberos)-
You will need the HDI privileged user keytab to consume from Kafka, the krb5.conf and jaas conf as well
Link to end to end sample
1. Keytab:
Generate the keytab for the privileged user authorized to consume from Kafka in your HDI cluster as follows, lets call the keytab, kafka-client-hdi.keytab-
ktutil
addent -password -p <UPN>@<REALM> -k 1 -e RC4-HMAC
wkt kafka-client-hdi.keytab
scp this to your local machine to copy onto the Kafka Connect worker Docker image.
2. krb5.conf:
The following is sample krb5.conf content; Create this on your local machine, modify to reflect your krb5 details; we will use this to copy onto the Kafka Connect worker Docker image-
[libdefaults]
default_realm = <yourKerberosRealm>
[realms]
<yourKerberosRealmName> = {
admin_server = <yourKerberosRealmNameInLowerCase>
kdc = <yourKerberosRealmNameInLowerCase>
default_domain = <yourDomainInLowerCase>
}
[domain_realm]
<yourKerberosRealmNameInLowerCase> = <yourKerberosRealm>
.<yourKerberosRealmNameInLowerCase> = <yourKerberosRealm>
[login]
krb4_convert = true
krb4_get_tickets = false
E.g.
Review this sample.
3. jaas.conf:
The following is sample jaas.conf content-
KafkaClient {
com.sun.security.auth.module.Krb5LoginModule required
useKeyTab=true
storeKey=true
keyTab="/etc/security/keytabs/kafka-client-hdi.keytab"
principal="<yourUserPrincipal>@<yourKerberosRealm>";
};
4. Configs to add to the Docker image:
This is covered in detail further on. It is specified here for the purpose of completeness of defining what goes onto the worker config.
COPY krb5.conf /etc/krb5.conf
COPY hdi-esp-jaas.conf /etc/hdi-esp-jaas.conf
COPY kafka-client-hdi.keytab /etc/security/keytabs/kafka-client-hdi.keytab
ENV KAFKA_OPTS="-Djava.security.krb5.conf=/etc/krb5.conf"
ENV CONNECT_CONNECTOR_CLIENT_CONFIG_OVERRIDE_POLICY=All
ENV CONNECT_SASL_MECHANISM=GSSAPI
ENV CONNECT_SASL_KERBEROS_SERVICE_NAME=kafka
ENV CONNECT_SECURITY_PROTOCOL=SASL_PLAINTEXT
ENV CONNECT_SASL_JAAS_CONFIG="com.sun.security.auth.module.Krb5LoginModule required useKeyTab=true storeKey=true keyTab=\"/etc/security/keytabs/kafka-client-hdi.keytab\" principal=\"<yourKerberosUPN>@<yourKerberosRealm>\";"
The following is complete set of connector sink properties-
| # | Property | Purpose | Details |
|---|---|---|---|
| 1 | connector.class | Classname of the Kusto sink | Hard code to com.microsoft.azure.kusto.kafka.connect.sink.KustoSinkConnectorRequired |
| 2 | topics | Kafka topic specification | List of topics separated by commas Required |
| 3 | kusto.ingestion.url | Kusto ingestion endpoint URL | Provide the ingest URL of your ADX cluster Use the following construct for the private URL - https://ingest-private-[cluster].kusto.windows.net Required |
| 4 | kusto.query.url | Kusto query endpoint URL | Provide the engine URL of your ADX cluster Optional |
| 5 | aad.auth.authority | Credentials for Kusto | Provide the tenant ID of your Azure Active Directory Required |
| 6 | aad.auth.appid | Credentials for Kusto | Provide Azure Active Directory Service Principal Name Required |
| 7 | aad.auth.appkey | Credentials for Kusto | Provide Azure Active Directory Service Principal secret Required |
| 8 | kusto.tables.topics.mapping | Mapping of topics to tables | Provide 1..many topic-table comma-separated mappings as follows- [{'topic': '<topicName1>','db': '<datebaseName>', 'table': '<tableName1>','format': '<format-e.g.avro/csv/json>', 'mapping':'<tableMappingName1>','streaming':'false'},{'topic': '<topicName2>','db': '<datebaseName>', 'table': '<tableName2>','format': '<format-e.g.avro/csv/json>', 'mapping':'<tableMappingName2>','streaming':'false'}] Required |
| 9 | key.converter | Deserialization | One of the below supported- org.apache.kafka.connect.storage.StringConverter org.apache.kafka.connect.json.JsonConverter io.confluent.connect.avro.AvroConverter io.confluent.connect.json.JsonSchemaConverter org.apache.kafka.connect.converters.ByteArrayConverter Required |
| 10 | value.converter | Deserialization | One of the below supported- org.apache.kafka.connect.storage.StringConverter org.apache.kafka.connect.json.JsonConverter io.confluent.connect.avro.AvroConverter io.confluent.connect.json.JsonSchemaConverter org.apache.kafka.connect.converters.ByteArrayConverter Required |
| 11 | value.converter.schema.registry.url | Schema validation | URI of the Kafka schema registry Optional |
| 12 | value.converter.schemas.enable | Schema validation | Set to true if you have embedded schema with payload but are not leveraging the schema registry Applicable for avro and json Optional |
| 13 | tasks.max | connector parallelism | Specify the number of connector copy/sink tasks Required |
| 14 | flush.size.bytes | Performance knob for batching | Maximum bufer byte size per topic+partition combination that in combination with flush.interval.ms (whichever is reached first) should result in sinking to Kusto Default - 1 MB Required |
| 15 | flush.interval.ms | Performance knob for batching | Minimum time interval per topic+partition combo that in combination with flush.size.bytes (whichever is reached first) should result in sinking to Kusto Default - 30 seconds Required |
| 16 | tempdir.path | Local directory path on Kafka Connect worker to buffer files to before shipping to Kusto | Default is value returned by System.getProperty("java.io.tmpdir") with a GUID attached to itOptional |
| 17 | behavior.on.error | Configurable behavior in response to errors encountered | Possible values - log, ignore, fail log - log the error, send record to dead letter queue, and continue processing ignore - log the error, send record to dead letter queue, proceed with processing despite errors encountered fail - shut down connector task upon encountering Default - fail Optional |
| 18 | errors.retry.max.time.ms | Configurable retries for transient errors | Period of time in milliseconds to retry for transient errors Default - 300 ms Optional |
| 19 | errors.retry.backoff.time.ms | Configurable retries for transient errors | Period of time in milliseconds to backoff before retry for transient errors Default - 10 ms Optional |
| 20 | errors.deadletterqueue.bootstrap.servers | Channel to write records that failed deserialization | CSV or kafkaBroker:port Optional |
| 21 | errors.deadletterqueue.topic.name | Channel to write records that failed deserialization | Pre-created topic name Optional |
| 22 | errors.deadletterqueue.security.protocol | Channel to write records that failed deserialization | Securitry protocol of secure Kafka cluster Optional but when feature is used with secure cluster, is required |
| 23 | errors.deadletterqueue.sasl.mechanism | Channel to write records that failed deserialization | SASL mechanism of secure Kafka cluster Optional but when feature is used with secure cluster, is required |
| 24 | errors.deadletterqueue.sasl.jaas.config | Channel to write records that failed deserialization | JAAS config of secure Kafka cluster Optional but when feature is used with secure cluster, is required |
| 25 | misc.deadletterqueue.bootstrap.servers | Channel to write records that due to reasons other than deserialization | CSV of kafkaBroker:port Optional |
| 26 | misc.deadletterqueue.topic.name | Channel to write records that due to reasons other than deserialization | Pre-created topic name Optional |
| 27 | misc.deadletterqueue.security.protocol | Channel to write records that due to reasons other than deserialization | Securitry protocol of secure Kafka cluster Optional but when feature is used with secure cluster, is required |
| 28 | misc.deadletterqueue.sasl.mechanism | Channel to write records that due to reasons other than deserialization | SASL mechanism of secure Kafka cluster Optional but when feature is used with secure cluster, is required |
| 29 | misc.deadletterqueue.sasl.jaas.config | Channel to write records that due to reasons other than deserialization | JAAS config of secure Kafka cluster Optional but when feature is used with secure cluster, is required |
| 30 | consumer.override.bootstrap.servers | Security details explicitly required for secure Kafka clusters | Bootstrap server:port CSV of secure Kafka cluster Required for secure Kafka clusters |
| 31 | consumer.override.security.protocol | Security details explicitly required for secure Kafka clusters | Security protocol of secure Kafka cluster Required for secure Kafka clusters |
| 32 | consumer.override.sasl.mechanism | Security details explicitly required for secure Kafka clusters | SASL mechanism of secure Kafka cluster Required for secure Kafka clusters |
| 33 | consumer.override.sasl.jaas.config | Security details explicitly required for secure Kafka clusters | JAAS config of secure Kafka cluster Required for secure Kafka clusters |
| 34 | consumer.override.sasl.kerberos.service.name | Security details explicitly required for secure Kafka clusters, specifically kerberized Kafka | Kerberos service name of kerberized Kafka cluster Required for kerberized Kafka clusters |
| 35 | consumer.override.auto.offset.reset | Configurable consuming from offset | Possible values are - earliest or latest Optional |
| 36 | consumer.override.max.poll.interval.ms | Config to prevent duplication | Set to a value to avoid consumer leaving the group while the Connector is retrying Optional |
Kusto supports Streaming ingestion in order to achieve sub-second latency.
This connector supports this using Managed streaming client.
Usage: configure per topic-table that streaming should be used. For example:
kusto.tables.topics.mapping=[{'topic': 't1','db': 'db', 'table': 't1','format': 'json', 'mapping':'map', 'streaming': true}].
Requirements: Streaming enabled on the cluster. Streaming policy configured on the table or database.
Additional configurations: flush.size.bytes and flush.interval.ms are still used to batch records together before ingestion - flush.size.bytes should not be over 4MB, flush.interval.ms is suggested to be low (hundreds of milliseconds). We still recommend configuring ingestion batching policy at the table or database level, as the client falls back to queued ingestion in case of failure and retry-exhaustion.
The following is the roadmap-
| # | Roadmap item |
|---|---|
| 1 | Schema evolution support |
Kafka Connect connectors can be deployed in standalone mode (just for development) or in distributed mode (production).
This involves having the connector plugin jar in /usr/share/java of a Kafka Connect worker, reference to the same plugin path in connect-standalone.properties, and launching of the connector from command line. This is not scalable, not fault tolerant, and is not recommeded for production.
Distributed Kafka Connect essentially involves creation of a KafkaConnect worker cluster as shown in the diagram below.
- Azure Kubernetes Service is a great infrastructure for the connect cluster, due to its managed and scalable nature
- Kubernetes is a great platform for the connect cluster, due to its scalable nature and self-healing
- Each orange polygon is a Kafka Connect worker and each green polygon is a sink connector instance
- A Kafka Connect worker can have 1..many task instances which helps with scale
- When a Kafka Connect worker is maxed out from a resource perspective (CPU, RAM), you can scale horizontally, add more Kafka Connect workers, ands tasks within them
- Kafka Connect service manages rebalancing of tasks to Kafka topic partitions automatically without pausing the connector tasks in recent versions of Kafka
- A Docker image needs to be created to deploy the Kusto sink connector in a distributed mode. This is detailed below.
Multiple options are available-
https://github.com/Azure/kafka-sink-azure-kusto/releases
https://www.confluent.io/hub/microsoftcorporation/kafka-sink-azure-kusto/
The dependencies are-
1. Clone the repo
git clone git://github.com/Azure/kafka-sink-azure-kusto.git
cd ./kafka-sink-azure-kusto2. Build with maven
For an Uber jar, run the below-
mvn clean compile assembly:singleFor the connector jar along with jars of associated dependencies, run the below-
mvn clean installLook within target/components/packages/microsoftcorporation-kafka-sink-azure-kusto-<version>/microsoftcorporation-kafka-sink-azure-kusto-<version>/lib/ folder
In a standalone mode (not recommended for production), the connector can be test driven in any of the following ways-
Review this hands on lab. It includes dockerized kafka, connector and Kafka producer to take away complexities and allow you to focus on the connector aspect.
Review this hands on lab. The lab referenced may have slightly outdated list of sink properties. Modify them to make current, leveraging the latest sink properties detailed in section 5.
The following are the components and configuration needed in place.
The following section strives to explain further, pictorially, what's involved with distributed deployment of the connector-
- Create a Docker Hub account if it does not exist
- Install Docker desktop on your machine
- Build a docker image for the KafkaConnect worker that include any connect worker level configurations, and the Kusto connector jar
- Push the image to the Docker hub
- Provision KafkaConnect workers on our Azure Kubernetes Service cluster
When we start off, all we have is an empty Kubernetes cluster-
When we are done, we have a live KafkaConnect cluster that is integrated with Confluent Cloud-
Note: This still does not have copy tasks (connector tasks) running yet
- Install Postman on our local machine
- Import KafkaConnect REST call JSON collection from Github into Postman
https://github.com/Azure/azure-kusto-labs/blob/confluent-clound-hol/kafka-integration/confluent-cloud/rest-calls/Confluent-Cloud-ADX-HoL-1-STUB.postman_collection.json
OR
8. Find the REST calls here to call from CLI
Note: Depending on Kafka security configuration, update the security configuration in the sink properties.
- Launch the Kafka-ADX copy tasks/REST call, otherwise called connector tasks from Postman or via curl command
This is what we will see, a Kusto sink connector cluster with copy tasks running.
Note: The diagram below depicts just one connector task per Kafka Connect worker. You can actually run 1..many connector tasks till you max out the capacity.
Section 11, below, links to a hands-on lab to test drive the deployment. The lab is end to end. Prefer the Confluent cloud lab for the simplest deployment.
The labs referenced below may have slightly outdated list of sink properties. Modify them to make current, leveraging the latest sink properties detailed in section 5.
For a non-secured HDInsight Kafka cluster - run through this end-to-end hands-on-lab
For a secured HDInsight Kafka cluster (Kerberised) - review these details specific to HDInsight with Enterprise Security Package
Run through this end-to-end hands-on-lab
Run through this end-to-end hands-on-lab
13. Apache Kafka version - Docker Kafka Connect base image version - Confluent Helm chart version related
We have deliberately not published a Docker image due to the multitude of versions of Apache Kafka bundled into the Confluent platform versions, and multiple versions of our Kafka connector, not to mention the security configs specific to each customer's Kafka deployment. We therefore recommend that a custom image be developed using the appropriate version compatible with the Apache Kafka version.
Similarly, we recommend leveraging the right version of the Helm chart. Confluent base helm chart for Kafka Connect.
The connector plugin is open source. We welcome feedback, and contribution. Log an issue, in the issues tab as needed. See section 14.
- Connector tasks can be scaled out per Kafka Connect worker by pausing, editing and resuming the connector cluster
- When tasks per worker are maxed out, Azure Kubernetes Service cluster can be scaled out for more nodes, and Kafka Connect workers can be provisioned on the same
- Confluent recommends Kafka Connect workers with minimum of 4 cores and 16 GB of RAM
- Start with 3 workers (3 AKS nodes), and scale horizontally as needed
- Number of tasks should ideally be equal to the number of Kafka topic partitions, not more
- Play with the number of tasks, wokers, nodes till you see the performance you desire
- Kafka topic: number of partitions should be tuned for performance
- Connectors: AKS right-sizing, connector tasks right-sizing, configure the right values for flush.size.bytes and flush.interval.ms
- Kusto: Right-size Kusto cluster for ingestion (SKU and node count), tune the table or database ingestion batching policy
- Format: Avro (with schema registry) and CSV perform more-or-less similarly from tests done
To upgrade, you would have to stop the connector tasks, recreate your connect worker Docker image to include the latest jar, update the sink properties to leverage the renamed and latest sink properties, reprovision the connect workers, then launch the copy tasks. You can use the consumer.override* feature to manipulate offset to read from.
- Found a bug? Please help us fix it by thoroughly documenting it and filing an issue.
- Have a feature request? Please post it on User Voice to help us prioritize
- Have a technical question? Ask on Stack Overflow with tag "azure-data-explorer"
- Need Support? Every customer with an active Azure subscription has access to support with guaranteed response time. Consider submitting a ticket and get assistance from Microsoft support team
With version 1.0, we overhauled the connector. The following are the changes-
- We renamed some properties for consistency with standards
- Added support for schema registry
- Added support for more converters - we supported only stringConverter and ByteArrayConverter previously
- Improved upfront validation and fail fast
- Added support for configurable behavior on error
- Added support for configurable retries
- Added support for Kafka Connect dead letter queues
- Introduced additional dead letter queue property for those errors that are not handled by Kafka Connect through its dead letter queue feature
- Improved the delivery guarantees to "at least once" (no data loss)
Here is our blog post.
To upgrade, you would have to stop the connector tasks, recreate your connect worker Docker image to include the latest jar, update the sink properties to leverage the renamed and latest sink properties, reprovision the connect workers, then launch the copy tasks.
For information about what changes are included in each release, please see the Release History section of this document.
| Release Version | Release Date | Changes Included |
|---|---|---|
| 0.1.0 | 2020-03-05 |
|
| 1.0.1 | 2020-08-04 |
|
| 1.0.2 | 2020-10-06 |
|
| 1.0.3 | 2020-10-13 |
|
| 2.0.0 | 2020-11-12 |
|
| 2.1.0 | 2021-07-11 |
|
| 2.2.0 | 2021-09-13 |
|
| 3.0.0 | 2022-06-06 |
|
| 3.0.1 | 2022-06-13 |
|
| 3.0.2 | 2022-07-20 |
|
This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.microsoft.com.
When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
In order to make the PR process efficient, please follow the below checklist:
- There is an issue open concerning the code added - Either a bug or enhancement. Preferably the issue includes an agreed upon approach.
- PR comment explains the changes done - This should be a TL;DR; as the rest of it should be documented in the related issue.
- PR is concise - Try to avoid making drastic changes in a single PR. Split it into multiple changes if possible. If you feel a major change is needed, make sure the commit history is clear and maintainers can comfortably review both the code and the logic behind the change.
- Please provide any related information needed to understand the change - Especially in the form of unit tests, but also docs, guidelines, use-case, best practices, etc as appropriate.
- Checks should pass
This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.