Amazon Kinesis ingestion

When you enable the Kinesis indexing service, you can configure supervisors on the Overlord to manage the creation and lifetime of Kinesis indexing tasks. These indexing tasks read events using Kinesis’ own shard and sequence number mechanism to guarantee exactly-once ingestion. The supervisor oversees the state of the indexing tasks to:

• coordinate handoffs
• manage failures
• ensure that scalability and replication requirements are maintained.

To use the Kinesis indexing service, load the druid-kinesis-indexing-service core Apache Druid extension (see Including Extensions).

Before you deploy the Kinesis extension to production, read the Kinesis known issues.

Submitting a Supervisor Spec

To use the Kinesis indexing service, load the druid-kinesis-indexing-service extension on both the Overlord and the MiddleManagers. Druid starts a supervisor for a dataSource when you submit a supervisor spec. Submit your supervisor spec to the following endpoint:

http://<OVERLORD_IP>:<OVERLORD_PORT>/druid/indexer/v1/supervisor

For example:

curl -X POST -H 'Content-Type: application/json' -d @supervisor-spec.json http://localhost:8090/druid/indexer/v1/supervisor

Where the file supervisor-spec.json contains a Kinesis supervisor spec:

{
  "type": "kinesis",
  "spec": {
    "dataSchema": {
      "dataSource": "metrics-kinesis",
      "timestampSpec": {
        "column": "timestamp",
        "format": "auto"
      },
     "dimensionsSpec": {
        "dimensions": [],
        "dimensionExclusions": [
         "timestamp",
         "value"
       ]
      },
     "metricsSpec": [
        {
         "name": "count",
          "type": "count"
       },
       {
          "name": "value_sum",
          "fieldName": "value",
          "type": "doubleSum"
        },
       {
         "name": "value_min",
         "fieldName": "value",
         "type": "doubleMin"
       },
        {
          "name": "value_max",
         "fieldName": "value",
         "type": "doubleMax"
       }
     ],
     "granularitySpec": {
        "type": "uniform",
        "segmentGranularity": "HOUR",
        "queryGranularity": "NONE"
     }
   },
    "ioConfig": {
     "stream": "metrics",
     "inputFormat": {
       "type": "json"
      },
      "endpoint": "kinesis.us-east-1.amazonaws.com",
      "taskCount": 1,
      "replicas": 1,
      "taskDuration": "PT1H",
     "recordsPerFetch": 2000,
      "fetchDelayMillis": 1000
   },
   "tuningConfig": {
     "type": "kinesis",
     "maxRowsPerSegment": 5000000
   }
  }
}

Supervisor Spec

KinesisSupervisorIOConfig

Task Autoscaler Properties

Note that Task AutoScaler is currently designated as experimental.

Lag Based AutoScaler Strategy Related Properties

The Kinesis indexing service reports lag metrics measured in time milliseconds rather than message count which is used by Kafka.

The following example demonstrates a supervisor spec with lagBased autoScaler enabled:

{
  "type": "kinesis",
  "dataSchema": {
    "dataSource": "metrics-kinesis",
    "timestampSpec": {
      "column": "timestamp",
      "format": "auto"
    },
    "dimensionsSpec": {
      "dimensions": [],
      "dimensionExclusions": [
        "timestamp",
        "value"
      ]
    },
    "metricsSpec": [
      {
        "name": "count",
        "type": "count"
      },
      {
        "name": "value_sum",
        "fieldName": "value",
        "type": "doubleSum"
      },
      {
        "name": "value_min",
        "fieldName": "value",
        "type": "doubleMin"
      },
      {
        "name": "value_max",
        "fieldName": "value",
        "type": "doubleMax"
      }
    ],
    "granularitySpec": {
      "type": "uniform",
      "segmentGranularity": "HOUR",
      "queryGranularity": "NONE"
    }
  },
  "ioConfig": {
    "stream": "metrics",
    "autoScalerConfig": {
      "enableTaskAutoScaler": true,
      "taskCountMax": 6,
      "taskCountMin": 2,
      "minTriggerScaleActionFrequencyMillis": 600000,
      "autoScalerStrategy": "lagBased",
      "lagCollectionIntervalMillis": 30000,
      "lagCollectionRangeMillis": 600000,
      "scaleOutThreshold": 600000,
      "triggerScaleOutFractionThreshold": 0.3,
      "scaleInThreshold": 100000,
      "triggerScaleInFractionThreshold": 0.9,
      "scaleActionStartDelayMillis": 300000,
      "scaleActionPeriodMillis": 60000,
      "scaleInStep": 1,
      "scaleOutStep": 2
    },
    "inputFormat": {
      "type": "json"
    },
    "endpoint": "kinesis.us-east-1.amazonaws.com",
    "taskCount": 1,
    "replicas": 1,
    "taskDuration": "PT1H",
    "recordsPerFetch": 2000,
    "fetchDelayMillis": 1000
  },
  "tuningConfig": {
    "type": "kinesis",
    "maxRowsPerSegment": 5000000
  }
}

Specifying data format

Kinesis indexing service supports both inputFormat and parser to specify the data format. The inputFormat is a new and recommended way to specify the data format for Kinesis indexing service, but unfortunately, it doesn’t support all data formats supported by the legacy parser. (They will be supported in the future.)

The supported inputFormats include csv, delimited, json, avro_stream, protobuf. You can also read thrift formats using parser.

KinesisSupervisorTuningConfig

The tuningConfig is optional and default parameters will be used if no tuningConfig is specified.

IndexSpec

Bitmap types

For Roaring bitmaps:

For Concise bitmaps:

SegmentWriteOutMediumFactory

Operations

This section gives descriptions of how some supervisor APIs work specifically in Kinesis Indexing Service. For all supervisor APIs, please check Supervisor APIs.

AWS Authentication

To authenticate with AWS, you must provide your AWS access key and AWS secret key via runtime.properties, for example:

-Ddruid.kinesis.accessKey=123 -Ddruid.kinesis.secretKey=456

The AWS access key ID and secret access key are used for Kinesis API requests. If this is not provided, the service will look for credentials set in environment variables, via Web Identity Token, in the default profile configuration file, and from the EC2 instance profile provider (in this order).

Getting Supervisor Status Report

GET /druid/indexer/v1/supervisor/<supervisorId>/status returns a snapshot report of the current state of the tasks managed by the given supervisor. This includes the latest sequence numbers as reported by Kinesis. Unlike the Kafka Indexing Service, Kinesis reports lag metrics measured in time difference in milliseconds between the current sequence number and latest sequence number, rather than message count.

The status report also contains the supervisor’s state and a list of recently thrown exceptions (reported as recentErrors, whose max size can be controlled using the druid.supervisor.maxStoredExceptionEvents configuration). There are two fields related to the supervisor’s state - state and detailedState. The state field will always be one of a small number of generic states that are applicable to any type of supervisor, while the detailedState field will contain a more descriptive, implementation-specific state that may provide more insight into the supervisor’s activities than the generic state field.

The list of possible state values are: [PENDING, RUNNING, SUSPENDED, STOPPING, UNHEALTHY_SUPERVISOR, UNHEALTHY_TASKS]

The list of detailedState values and their corresponding state mapping is as follows:

On each iteration of the supervisor’s run loop, the supervisor completes the following tasks in sequence:

1. Fetch the list of shards from Kinesis and determine the starting sequence number for each shard (either based on the last processed sequence number if continuing, or starting from the beginning or ending of the stream if this is a new stream).
2. Discover any running indexing tasks that are writing to the supervisor’s datasource and adopt them if they match the supervisor’s configuration, else signal them to stop.
3. Send a status request to each supervised task to update our view of the state of the tasks under our supervision.
4. Handle tasks that have exceeded taskDuration and should transition from the reading to publishing state.
5. Handle tasks that have finished publishing and signal redundant replica tasks to stop.
6. Handle tasks that have failed and clean up the supervisor’s internal state.
7. Compare the list of healthy tasks to the requested taskCount and replicas configurations and create additional tasks if required.

The detailedState field will show additional values (those marked with “first iteration only”) the first time the supervisor executes this run loop after startup or after resuming from a suspension. This is intended to surface initialization-type issues, where the supervisor is unable to reach a stable state (perhaps because it can’t connect to Kinesis, it can’t read from the stream, or it can’t communicate with existing tasks). Once the supervisor is stable - that is, once it has completed a full execution without encountering any issues - detailedState will show a RUNNING state until it is stopped, suspended, or hits a failure threshold and transitions to an unhealthy state.

Updating Existing Supervisors

POST /druid/indexer/v1/supervisor can be used to update existing supervisor spec. Calling this endpoint when there is already an existing supervisor for the same dataSource will cause:

• The running supervisor to signal its managed tasks to stop reading and begin publishing.
• The running supervisor to exit.
• A new supervisor to be created using the configuration provided in the request body. This supervisor will retain the existing publishing tasks and will create new tasks starting at the sequence numbers the publishing tasks ended on.

Seamless schema migrations can thus be achieved by simply submitting the new schema using this endpoint.

Suspending and Resuming Supervisors

You can suspend and resume a supervisor using POST /druid/indexer/v1/supervisor/<supervisorId>/suspend and POST /druid/indexer/v1/supervisor/<supervisorId>/resume, respectively.

Note that the supervisor itself will still be operating and emitting logs and metrics, it will just ensure that no indexing tasks are running until the supervisor is resumed.

Resetting Supervisors

The POST /druid/indexer/v1/supervisor/<supervisorId>/reset operation clears stored sequence numbers, causing the supervisor to start reading from either the earliest or latest sequence numbers in Kinesis (depending on the value of useEarliestSequenceNumber). After clearing stored sequence numbers, the supervisor kills and recreates active tasks, so that tasks begin reading from valid sequence numbers.

Use care when using this operation! Resetting the supervisor may cause Kinesis messages to be skipped or read twice, resulting in missing or duplicate data.

The reason for using this operation is to recover from a state in which the supervisor ceases operating due to missing sequence numbers. The indexing service keeps track of the latest persisted sequence number in order to provide exactly-once ingestion guarantees across tasks.

Subsequent tasks must start reading from where the previous task completed in order for the generated segments to be accepted. If the messages at the expected starting sequence numbers are no longer available in Kinesis (typically because the message retention period has elapsed or the topic was removed and re-created) the supervisor will refuse to start and in-flight tasks will fail. This operation enables you to recover from this condition.

Note that the supervisor must be running for this endpoint to be available.

Terminating Supervisors

The POST /druid/indexer/v1/supervisor/<supervisorId>/terminate operation terminates a supervisor and causes all associated indexing tasks managed by this supervisor to immediately stop and begin publishing their segments. This supervisor will still exist in the metadata store and its history may be retrieved with the supervisor history API, but will not be listed in the ‘get supervisors’ API response nor can its configuration or status report be retrieved. The only way this supervisor can start again is by submitting a functioning supervisor spec to the create API.

Capacity Planning

Kinesis indexing tasks run on MiddleManagers and are thus limited by the resources available in the MiddleManager cluster. In particular, you should make sure that you have sufficient worker capacity (configured using the druid.worker.capacity property) to handle the configuration in the supervisor spec. Note that worker capacity is shared across all types of indexing tasks, so you should plan your worker capacity to handle your total indexing load (e.g. batch processing, realtime tasks, merging tasks, etc.). If your workers run out of capacity, Kinesis indexing tasks will queue and wait for the next available worker. This may cause queries to return partial results but will not result in data loss (assuming the tasks run before Kinesis purges those sequence numbers).

A running task will normally be in one of two states: reading or publishing. A task will remain in reading state for taskDuration, at which point it will transition to publishing state. A task will remain in publishing state for as long as it takes to generate segments, push segments to deep storage, and have them be loaded and served by a Historical process (or until completionTimeout elapses).

The number of reading tasks is controlled by replicas and taskCount. In general, there will be replicas * taskCount reading tasks, the exception being if taskCount > {numKinesisShards} in which case {numKinesisShards} tasks will be used instead. When taskDuration elapses, these tasks will transition to publishing state and replicas * taskCount new reading tasks will be created. Therefore to allow for reading tasks and publishing tasks to run concurrently, there should be a minimum capacity of:

workerCapacity = 2 * replicas * taskCount

This value is for the ideal situation in which there is at most one set of tasks publishing while another set is reading. In some circumstances, it is possible to have multiple sets of tasks publishing simultaneously. This would happen if the time-to-publish (generate segment, push to deep storage, loaded on Historical) > taskDuration. This is a valid scenario (correctness-wise) but requires additional worker capacity to support. In general, it is a good idea to have taskDuration be large enough that the previous set of tasks finishes publishing before the current set begins.

Supervisor Persistence

When a supervisor spec is submitted via the POST /druid/indexer/v1/supervisor endpoint, it is persisted in the configured metadata database. There can only be a single supervisor per dataSource, and submitting a second spec for the same dataSource will overwrite the previous one.

When an Overlord gains leadership, either by being started or as a result of another Overlord failing, it will spawn a supervisor for each supervisor spec in the metadata database. The supervisor will then discover running Kinesis indexing tasks and will attempt to adopt them if they are compatible with the supervisor’s configuration. If they are not compatible because they have a different ingestion spec or shard allocation, the tasks will be killed and the supervisor will create a new set of tasks. In this way, the supervisors are persistent across Overlord restarts and fail-overs.

A supervisor is stopped via the POST /druid/indexer/v1/supervisor/<supervisorId>/terminate endpoint. This places a tombstone marker in the database (to prevent the supervisor from being reloaded on a restart) and then gracefully shuts down the currently running supervisor. When a supervisor is shut down in this way, it will instruct its managed tasks to stop reading and begin publishing their segments immediately. The call to the shutdown endpoint will return after all tasks have been signalled to stop but before the tasks finish publishing their segments.

Schema/Configuration Changes

Schema and configuration changes are handled by submitting the new supervisor spec via the same POST /druid/indexer/v1/supervisor endpoint used to initially create the supervisor. The Overlord will initiate a graceful shutdown of the existing supervisor which will cause the tasks being managed by that supervisor to stop reading and begin publishing their segments. A new supervisor will then be started which will create a new set of tasks that will start reading from the sequence numbers where the previous now-publishing tasks left off, but using the updated schema. In this way, configuration changes can be applied without requiring any pause in ingestion.

Deployment Notes

On the Subject of Segments

Each Kinesis Indexing Task puts events consumed from Kinesis Shards assigned to it in a single segment for each segment granular interval until maxRowsPerSegment, maxTotalRows or intermediateHandoffPeriod limit is reached, at this point a new shard for this segment granularity is created for further events. Kinesis Indexing Task also does incremental hand-offs which means that all the segments created by a task will not be held up till the task duration is over. As soon as maxRowsPerSegment, maxTotalRows or intermediateHandoffPeriod limit is hit, all the segments held by the task at that point in time will be handed-off and new set of segments will be created for further events. This means that the task can run for longer durations of time without accumulating old segments locally on Middle Manager processes and it is encouraged to do so.

Kinesis Indexing Service may still produce some small segments. Lets say the task duration is 4 hours, segment granularity is set to an HOUR and Supervisor was started at 9:10 then after 4 hours at 13:10, new set of tasks will be started and events for the interval 13:00 - 14:00 may be split across previous and new set of tasks. If you see it becoming a problem then one can schedule re-indexing tasks be run to merge segments together into new segments of an ideal size (in the range of ~500-700 MB per segment). Details on how to optimize the segment size can be found on Segment size optimization. There is also ongoing work to support automatic segment compaction of sharded segments as well as compaction not requiring Hadoop (see here).

Determining Fetch Settings

Internally, the Kinesis Indexing Service uses the Kinesis Record Supplier abstraction for fetching Kinesis data records and storing the records locally. The way the Kinesis Record Supplier fetches records is to have a separate thread run the fetching operation per each Kinesis Shard, the max number of threads is determined by fetchThreads. For example, a Kinesis stream with 3 shards will have 3 threads, each fetching from a shard separately. There is a delay between each fetching operation, which is controlled by fetchDelayMillis. The maximum number of records to be fetched per thread per operation is controlled by recordsPerFetch. Note that this is not the same as maxRecordsPerPoll.

The records fetched by each thread will be pushed to a queue in the order that they are fetched. The records are stored in this queue until poll() is called by either the supervisor or the indexing task. poll() will attempt to drain the internal buffer queue up to a limit of max(maxRecordsPerPoll, q.size()). Here maxRecordsPerPoll controls the theoretical maximum records to drain out of the buffer queue, so setting this parameter to a reasonable value is essential in preventing the queue from overflowing or memory exceeding heap size.

Kinesis places the following restrictions on calls to fetch records:

• Each data record can be up to 1 MB in size.
• Each shard can support up to 5 transactions per second for reads.
• Each shard can read up to 2 MB per second.
• The maximum size of data that GetRecords can return is 10 MB.

Values for recordsPerFetch and fetchDelayMillis should be chosen to maximize throughput under the above constraints. The values that you choose will depend on the average size of a record and the number of consumers you have reading from a given shard (which will be replicas unless you have other consumers also reading from this Kinesis stream).

If the above limits are violated, AWS will throw ProvisionedThroughputExceededException errors on subsequent calls to read data. When this happens, the Kinesis indexing service will pause by fetchDelayMillis and then attempt the call again.

Internally, each indexing task maintains a buffer that stores the fetched but not yet processed record. recordsPerFetch and fetchDelayMillis control this behavior. The number of records that the indexing task fetch from the buffer is controlled by maxRecordsPerPoll, which determines the number of records to be processed per each ingestion loop in the task.

Deaggregation

See issue

The Kinesis indexing service supports de-aggregation of multiple rows packed into a single record by the Kinesis Producer Library’s aggregate method for more efficient data transfer. Currently, enabling the de-aggregate functionality requires the user to manually provide the Kinesis Client Library on the classpath, since this library has a license not compatible with Apache projects.

To enable this feature, add the amazon-kinesis-client (tested on version 1.9.2) jar file (link) under dist/druid/extensions/druid-kinesis-indexing-service/. Then when submitting a supervisor-spec, set deaggregate to true.

Resharding

When changing the shard count for a Kinesis stream, there will be a window of time around the resharding operation with early shutdown of Kinesis ingestion tasks and possible task failures.

The early shutdowns and task failures are expected, and they occur because the supervisor will update the shard -> task group mappings as shards are closed and fully read, to ensure that tasks are not running with an assignment of closed shards that have been fully read and to ensure a balanced distribution of active shards across tasks.

This window with early task shutdowns and possible task failures will conclude when:

• All closed shards have been fully read and the Kinesis ingestion tasks have published the data from those shards, committing the “closed” state to metadata storage
• Any remaining tasks that had inactive shards in the assignment have been shutdown (these tasks would have been created before the closed shards were completely drained)

Kinesis known issues

Before you deploy the Kinesis extension to production, consider the following known issues:

• Avoid implementing more than one Kinesis supervisor that read from the same Kinesis stream for ingestion. Kinesis has a per-shard read throughput limit and having multiple supervisors on the same stream can reduce available read throughput for an individual Supervisor’s tasks. Additionally, multiple Supervisors ingesting to the same Druid Datasource can cause increased contention for locks on the Datasource.
• The only way to change the stream reset policy is to submit a new ingestion spec and set up a new supervisor.
• Timeouts for retrieving earliest sequence number will cause a reset of the supervisor. The job will resume own its own eventually, but it can trigger alerts.
• The Kinesis supervisor will not make progress if you have empty shards. Make sure you have at least 1 record in the shard.
• If ingestion tasks get stuck, the supervisor does not automatically recover. You should monitor ingestion tasks and investigate if your ingestion falls behind.
• A Kinesis supervisor can sometimes compare the checkpoint offset to retention window of the stream to see if it has fallen behind. These checks fetch the earliest sequence number for Kinesis which can result in IteratorAgeMilliseconds becoming very high in AWS CloudWatch.