Skip to content

Latest commit

 

History

History
1200 lines (834 loc) · 47.4 KB

INTRODUCTION.md

File metadata and controls

1200 lines (834 loc) · 47.4 KB

//@file INTRODUCTION.md

Introduction to librdkafka - the Apache Kafka C/C++ client library

librdkafka is a high performance C implementation of the Apache Kafka client, providing a reliable and performant client for production use. librdkafka also provides a native C++ interface.

Contents

The following chapters are available in this document

Performance

librdkafka is a multi-threaded library designed for use on modern hardware and it attempts to keep memory copying at a minimal. The payload of produced or consumed messages may pass through without any copying (if so desired by the application) putting no limit on message sizes.

librdkafka allows you to decide if high throughput is the name of the game, or if a low latency service is required, all through the configuration property interface.

The two most important configuration properties for performance tuning are:

  • batch.num.messages - the maximum number of messages to wait for to accumulate in the local queue before sending off a message set.
  • queue.buffering.max.ms - how long to wait for batch.num.messages to fill up in the local queue. A lower value improves latency at the cost of lower throughput and higher per-message overhead. A higher value improves throughput at the expense of latency. The recommended value for high throughput is > 50ms.

Performance numbers

The following performance numbers stem from tests using the following setup:

  • Intel Quad Core i7 at 3.4GHz, 8GB of memory
  • Disk performance has been shortcut by setting the brokers' flush configuration properties as so:
    • log.flush.interval.messages=10000000
    • log.flush.interval.ms=100000
  • Two brokers running on the same machine as librdkafka.
  • One topic with two partitions.
  • Each broker is leader for one partition each.
  • Using rdkafka_performance program available in the examples subdir.

Test results

  • Test1: 2 brokers, 2 partitions, required.acks=2, 100 byte messages: 850000 messages/second, 85 MB/second

  • Test2: 1 broker, 1 partition, required.acks=0, 100 byte messages: 710000 messages/second, 71 MB/second

  • Test3: 2 broker2, 2 partitions, required.acks=2, 100 byte messages, snappy compression: 300000 messages/second, 30 MB/second

  • Test4: 2 broker2, 2 partitions, required.acks=2, 100 byte messages, gzip compression: 230000 messages/second, 23 MB/second

Note: See the Test details chapter at the end of this document for information about the commands executed, etc.

Note: Consumer performance tests will be announced soon.

High throughput

The key to high throughput is message batching - waiting for a certain amount of messages to accumulate in the local queue before sending them off in one large message set or batch to the peer. This amortizes the messaging overhead and eliminates the adverse effect of the round trip time (rtt).

queue.buffering.max.ms (also called linger.ms) allows librdkafka to wait up to the specified amount of time to accumulate up to batch.num.messages in a single batch (MessageSet) before sending to the broker. The larger the batch the higher the throughput. Enabling msg debugging (set debug property to msg) will emit log messages for the accumulation process which lets you see what batch sizes are being produced.

Example using queue.buffering.max.ms=1:

... test [0]: MessageSet with 1514 message(s) delivered
... test [3]: MessageSet with 1690 message(s) delivered
... test [0]: MessageSet with 1720 message(s) delivered
... test [3]: MessageSet with 2 message(s) delivered
... test [3]: MessageSet with 4 message(s) delivered
... test [0]: MessageSet with 4 message(s) delivered
... test [3]: MessageSet with 11 message(s) delivered

Example using queue.buffering.max.ms=1000:

... test [0]: MessageSet with 10000 message(s) delivered
... test [0]: MessageSet with 10000 message(s) delivered
... test [0]: MessageSet with 4667 message(s) delivered
... test [3]: MessageSet with 10000 message(s) delivered
... test [3]: MessageSet with 10000 message(s) delivered
... test [3]: MessageSet with 4476 message(s) delivered

The default setting of queue.buffering.max.ms=1 is not suitable for high throughput, it is recommended to set this value to >50ms, with throughput leveling out somewhere around 100-1000ms depending on message produce pattern and sizes.

These setting are set globally (rd_kafka_conf_t) but applies on a per topic+partition basis.

Low latency

When low latency messaging is required the queue.buffering.max.ms should be tuned to the maximum permitted producer-side latency. Setting queue.buffering.max.ms to 1 will make sure messages are sent as soon as possible. You could check out How to decrease message latency to find more details. Lower buffering time leads to smaller batches and larger per-message overheads, increasing network, memory and CPU usage for producers, brokers and consumers.

Latency measurement

End-to-end latency is preferably measured by synchronizing clocks on producers and consumers and using the message timestamp on the consumer to calculate the full latency. Make sure the topic's log.message.timestamp.type is set to the default CreateTime (Kafka topic configuration, not librdkafka topic).

Latencies are typically incurred by the producer, network and broker, the consumer effect on end-to-end latency is minimal.

To break down the end-to-end latencies and find where latencies are adding up there are a number of metrics available through librdkafka statistics on the producer:

  • brokers[].int_latency is the time, per message, between produce() and the message being written to a MessageSet and ProduceRequest. High int_latency indicates CPU core contention: check CPU load and, involuntary context switches (/proc/<..>/status). Consider using a machine/instance with more CPU cores. This metric is only relevant on the producer.

  • brokers[].outbuf_latency is the time, per protocol request (such as ProduceRequest), between the request being enqueued (which happens right after int_latency) and the time the request is written to the TCP socket connected to the broker. High outbuf_latency indicates CPU core contention or network congestion: check CPU load and socket SendQ (netstat -anp | grep :9092).

  • brokers[].rtt is the time, per protocol request, between the request being written to the TCP socket and the time the response is received from the broker. High rtt indicates broker load or network congestion: check broker metrics, local socket SendQ, network performance, etc.

  • brokers[].throttle is the time, per throttled protocol request, the broker throttled/delayed handling of a request due to usage quotas. The throttle time will also be reflected in rtt.

  • topics[].batchsize is the size of individual Producer MessageSet batches. See below.

  • topics[].batchcnt is the number of messages in individual Producer MessageSet batches. Due to Kafka protocol overhead a batch with few messages will have a higher relative processing and size overhead than a batch with many messages. Use the linger.ms client configuration property to set the maximum amount of time allowed for accumulating a single batch, the larger the value the larger the batches will grow, thus increasing efficiency. When producing messages at a high rate it is recommended to increase linger.ms, which will improve throughput and in some cases also latency.

See STATISTICS.md for the full definition of metrics. A JSON schema for the statistics is available in statistics-schema.json.

Compression

Producer message compression is enabled through the compression.codec configuration property.

Compression is performed on the batch of messages in the local queue, the larger the batch the higher likelyhood of a higher compression ratio. The local batch queue size is controlled through the batch.num.messages and queue.buffering.max.ms configuration properties as described in the High throughput chapter above.

Message reliability

Message reliability is an important factor of librdkafka - an application can rely fully on librdkafka to deliver a message according to the specified configuration (request.required.acks and message.send.max.retries, etc).

If the topic configuration property request.required.acks is set to wait for message commit acknowledgements from brokers (any value but 0, see CONFIGURATION.md for specifics) then librdkafka will hold on to the message until all expected acks have been received, gracefully handling the following events:

  • Broker connection failure
  • Topic leader change
  • Produce errors signaled by the broker
  • Network problems

We recommend request.required.acks to be set to all to make sure produced messages are acknowledged by all in-sync replica brokers.

This is handled automatically by librdkafka and the application does not need to take any action at any of the above events. The message will be resent up to message.send.max.retries times before reporting a failure back to the application.

The delivery report callback is used by librdkafka to signal the status of a message back to the application, it will be called once for each message to report the status of message delivery:

  • If error_code is non-zero the message delivery failed and the error_code indicates the nature of the failure (rd_kafka_resp_err_t enum).
  • If error_code is zero the message has been successfully delivered.

See Producer API chapter for more details on delivery report callback usage.

The delivery report callback is optional but highly recommended.

Producer message delivery success

When a ProduceRequest is successfully handled by the broker and a ProduceResponse is received (also called the ack) without an error code the messages from the ProduceRequest are enqueued on the delivery report queue (if a delivery report callback has been set) and will be passed to the application on the next invocation rd_kafka_poll().

Producer message delivery failure

The following sub-chapters explains how different produce errors are handled.

If the error is retryable and there are remaining retry attempts for the given message(s), an automatic retry will be scheduled by librdkafka, these retries are not visible to the application.

Only permanent errors and temporary errors that have reached their maximum retry count will generate a delivery report event to the application with an error code set.

The application should typically not attempt to retry producing the message on failure, but instead configure librdkafka to perform these retries using the retries and retry.backoff.ms configuration properties.

Error: Timed out in transmission queue

Internal error ERR__TIMED_OUT_QUEUE.

The connectivity to the broker may be stalled due to networking contention, local or remote system issues, etc, and the request has not yet been sent.

The producer can be certain that the message has not been sent to the broker.

This is a retryable error, but is not counted as a retry attempt since the message was never actually transmitted.

A retry by librdkafka at this point will not cause duplicate messages.

Error: Timed out in flight to/from broker

Internal error ERR__TIMED_OUT, ERR__TRANSPORT.

Same reasons as for Timed out in transmission queue above, with the difference that the message may have been sent to the broker and might be stalling waiting for broker replicas to ack the message, or the response could be stalled due to networking issues. At this point the producer can't know if the message reached the broker, nor if the broker wrote the message to disk and replicas.

This is a retryable error.

A retry by librdkafka at this point may cause duplicate messages.

Error: Temporary broker-side error

Broker errors ERR_REQUEST_TIMED_OUT, ERR_NOT_ENOUGH_REPLICAS, ERR_NOT_ENOUGH_REPLICAS_AFTER_APPEND.

These errors are considered temporary and librdkafka is will retry them if permitted by configuration.

Error: Temporary errors due to stale metadata

Broker errors ERR_LEADER_NOT_AVAILABLE, ERR_NOT_LEADER_FOR_PARTITION.

These errors are considered temporary and a retry is warranted, a metadata request is automatically sent to find a new leader for the partition.

A retry by librdkafka at this point will not cause duplicate messages.

Error: Local time out

Internal error ERR__MSG_TIMED_OUT.

The message could not be successfully transmitted before message.timeout.ms expired, typically due to no leader being available or no broker connection. The message may have been retried due to other errors but those error messages are abstracted by the ERR__MSG_TIMED_OUT error code.

Since the message.timeout.ms has passed there will be no more retries by librdkafka.

Error: Permanent errors

Any other error is considered a permanent error and the message will fail immediately, generating a delivery report event with the distinctive error code.

The full list of permanent errors depend on the broker version and will likely grow in the future.

Typical permanent broker errors are:

  • ERR_CORRUPT_MESSAGE
  • ERR_MSG_SIZE_TOO_LARGE - adjust client's or broker's message.max.bytes.
  • ERR_UNKNOWN_TOPIC_OR_PART - topic or partition does not exist, automatic topic creation is disabled on the broker or the application is specifying a partition that does not exist.
  • ERR_RECORD_LIST_TOO_LARGE
  • ERR_INVALID_REQUIRED_ACKS
  • ERR_TOPIC_AUTHORIZATION_FAILED
  • ERR_UNSUPPORTED_FOR_MESSAGE_FORMAT
  • ERR_CLUSTER_AUTHORIZATION_FAILED

Producer retries

The ProduceRequest itself is not retried, instead the messages are put back on the internal partition queue by an insert sort that maintains their original position (the message order is defined at the time a message is initially appended to a partition queue, i.e., after partitioning). A backoff time (retry.backoff.ms) is set on the retried messages which effectively blocks retry attempts until the backoff time has expired.

Reordering

As for all retries, if max.in.flight > 1 and retries > 0, retried messages may be produced out of order, since a sub-sequent message in a sub-sequent ProduceRequest may already be in-flight (and accepted by the broker) by the time the retry for the failing message is sent.

Using the Idempotent Producer prevents reordering even with max.in.flight > 1, see Idempotent Producer below for more information.

Idempotent Producer

librdkafka supports the idempotent producer which provides strict ordering and and exactly-once producer guarantees. The idempotent producer is enabled by setting the enable.idempotence configuration property to true, this will automatically adjust a number of other configuration properties to adhere to the idempotency requirements, see the documentation of enable.idempotence in [CONFIGURATION.md] for more information. Producer instantiation will fail if the user supplied an incompatible value for any of the automatically adjusted properties, e.g., it is an error to explicitly set acks=1 when enable.idempotence=true is set.

Guarantees

There are three types of guarantees that the idempotent producer can satisfy:

  • Exactly-once - a message is only written to the log once. Does NOT cover the exactly-once consumer case.
  • Ordering - a series of messages are written to the log in the order they were produced.
  • Gap-less - a series of messages are written once and in order without risk of skipping messages. The sequence of messages may be cut short and fail before all messages are written, but may not fail individual messages in the series. This guarantee is enabled by default, but may be disabled by setting enable.gapless.guarantee if individual message failure is not a concern.

All three guarantees are in effect when idempotence is enabled, only gap-less may be disabled individually.

Ordering and message sequence numbers

librdkafka maintains the original produce() ordering per-partition for all messages produced, using an internal per-partition 64-bit counter called the msgid which starts at 1. This msgid allows messages to be re-inserted in the partition message queue in the original order in the case of retries.

The Idempotent Producer functionality in the Kafka protocol also has a per-message sequence number, which is a signed 32-bit wrapping counter that is reset each time the Producer's ID (PID) or Epoch changes.

The librdkafka msgid is used, along with a base msgid value stored at the time the PID/Epoch was bumped, to calculate the Kafka protocol's message sequence number.

With Idempotent Producer enabled there is no risk of reordering despite max.in.flight > 1 (capped at 5).

Note: "MsgId" in log messages refer to the librdkafka msgid, while "seq" refers to the protocol message sequence, "baseseq" is the seq of the first message in a batch. MsgId starts at 1, while message seqs start at 0.

The producer statistics also maintain two metrics for tracking the next expected response sequence:

  • next_ack_seq - the next sequence to expect an acknowledgement for, which is the last successfully produced MessageSet's last sequence + 1.
  • next_err_seq - the next sequence to expect an error for, which is typically the same as next_ack_seq until an error occurs, in which case the next_ack_seq can't be incremented (since no messages were acked on error). next_err_seq is used to properly handle sub-sequent errors due to a failing first request.

Note: Both are exposed in partition statistics.

Partitioner considerations

Strict ordering is guaranteed on a per partition basis.

An application utilizing the idempotent producer should not mix producing to explicit partitions with partitioner-based partitions since messages produced for the latter are queued separately until a topic's partition count is known, which would insert these messages after the partition-explicit messages regardless of produce order.

Leader change

There are corner cases where an Idempotent Producer has outstanding ProduceRequests in-flight to the previous leader while a new leader is elected.

A leader change is typically triggered by the original leader failing or terminating, which has the risk of also failing (some of) the in-flight ProduceRequests to that broker. To recover the producer to a consistent state it will not send any ProduceRequests for these partitions to the new leader broker until all responses for any outstanding ProduceRequests to the previous partition leader has been received, or these requests have timed out. This drain may take up to min(socket.timeout.ms, message.timeout.ms). If the connection to the previous broker goes down the outstanding requests are failed immediately.

Error handling

Background: The error handling for the Idempotent Producer, as initially proposed in the EOS design document, missed some corner cases which are now being addressed in KIP-360. There were some intermediate fixes and workarounds prior to KIP-360 that proved to be incomplete and made the error handling in the client overly complex. With the benefit of hindsight the librdkafka implementation will attempt to provide correctness from the lessons learned in the Java client and provide stricter and less complex error handling.

Note: At the time of this writing KIP-360 has not been accepted.

The follow sections describe librdkafka's handling of the Idempotent Producer specific errors that may be returned by the broker.

RD_KAFKA_RESP_ERR_OUT_OF_ORDER_SEQUENCE_NUMBER

This error is returned by the broker when the sequence number in the ProduceRequest is larger than the expected next sequence for the given PID+Epoch+Partition (last BaseSeq + msgcount + 1). Note: sequence 0 is always accepted.

If the failed request is the head-of-line (next expected sequence to be acked) it indicates desynchronization between the client and broker: the client thinks the sequence number is correct but the broker disagrees. There is no way for the client to recover from this scenario without risking message loss or duplication, and it is not safe for the application to manually retry messages. A fatal error (RD_KAFKA_RESP_ERR_OUT_OF_ORDER_SEQUENCE_NUMBER) is raised.

When the request is not head-of-line the previous request failed (for any reason), which means the messages in the current request can be retried after waiting for all outstanding requests for this partition to drain and then reset the Producer ID and start over.

Java Producer behaviour: Fail the batch, reset the pid, and then continue producing (and retrying sub-sequent) messages. This will lead to gaps in the message series.

RD_KAFKA_RESP_ERR_DUPLICATE_SEQUENCE_NUMBER

Returned by broker when the request's base sequence number is less than the expected sequence number (which is the last written sequence + msgcount). Note: sequence 0 is always accepted.

This error is typically benign and occurs upon retrying a previously successful send that was not acknowledged.

The messages will be considered successfully produced but will have neither timestamp or offset set.

Java Producer behaviour: Treats the message as successfully delivered.

RD_KAFKA_RESP_ERR_UNKNOWN_PRODUCER_ID

Returned by broker when the PID+Epoch is unknown, which may occur when the PID's state has expired (due to topic retention, DeleteRercords, or compaction).

The Java producer added quite a bit of error handling for this case, extending the ProduceRequest protocol to return the logStartOffset to give the producer a chance to differentiate between an actual UNKNOWN_PRODUCER_ID or topic retention having deleted the last message for this producer (effectively voiding the Producer ID cache). This workaround proved to be error prone (see explanation in KIP-360) when the partition leader changed.

KIP-360 suggests removing this error checking in favour of failing fast, librdkafka follows suite.

If the response is for the first ProduceRequest in-flight and there are no messages waiting to be retried nor any ProduceRequests unaccounted for, then the error is ignored and the epoch is incremented, this is likely to happen for an idle producer who's last written message has been deleted from the log, and thus its PID state. Otherwise the producer raises a fatal error (RD_KAFKA_RESP_ERR_UNKNOWN_PRODUCER_ID) since the delivery guarantees can't be satisfied.

Java Producer behaviour: Retries the send in some cases (but KIP-360 will change this). Not a fatal error in any case.

Standard errors

All the standard Produce errors are handled in the usual way, permanent errors will fail the messages in the batch, while temporary errors will be retried (if retry count permits).

If a permanent error is returned for a batch in a series of in-flight batches, the sub-sequent batches will fail with RD_KAFKA_RESP_ERR_OUT_OF_ORDER_SEQUENCE_NUMBER since the sequence number of the failed batched was never written to the topic log and next expected sequence thus not incremented on the broker.

A fatal error (RD_KAFKA_RESP_ERR__GAPLESS_GUARANTEE) is raised to satisfy the gap-less guarantee (if enable.gapless.guarantee is set) by failing all queued messages.

Message persistance status

To help the application decide what to do in these error cases, a new per-message API is introduced, rd_kafka_message_persistance_status(), which returns one of the following values:

  • RD_KAFKA_MSG_STATUS_NOT_PERSISTED - the message has never been transmitted to the broker, or failed with an error indicating it was not written to the log. Application retry will risk ordering, but not duplication.
  • RD_KAFKA_MSG_STATUS_POSSIBLY_PERSISTED - the message was transmitted to the broker, but no acknowledgement was received. Application retry will risk ordering and duplication.
  • RD_KAFKA_MSG_STATUS_PERSISTED - the message was written to the log by the broker and fully acknowledged. No reason for application to retry.

This method should be called by the application on delivery report error.

Usage

Documentation

The librdkafka API is documented in the rdkafka.h header file, the configuration properties are documented in CONFIGURATION.md

Initialization

The application needs to instantiate a top-level object rd_kafka_t which is the base container, providing global configuration and shared state. It is created by calling rd_kafka_new().

It also needs to instantiate one or more topics (rd_kafka_topic_t) to be used for producing to or consuming from. The topic object holds topic-specific configuration and will be internally populated with a mapping of all available partitions and their leader brokers. It is created by calling rd_kafka_topic_new().

Both rd_kafka_t and rd_kafka_topic_t comes with a configuration API which is optional. Not using the API will cause librdkafka to use its default values which are documented in CONFIGURATION.md.

Note: An application may create multiple rd_kafka_t objects and they share no state.

Note: An rd_kafka_topic_t object may only be used with the rd_kafka_t object it was created from.

Configuration

To ease integration with the official Apache Kafka software and lower the learning curve, librdkafka implements identical configuration properties as found in the official clients of Apache Kafka.

Configuration is applied prior to object creation using the rd_kafka_conf_set() and rd_kafka_topic_conf_set() APIs.

Note: The rd_kafka.._conf_t objects are not reusable after they have been passed to rd_kafka.._new(). The application does not need to free any config resources after a rd_kafka.._new() call.

Example

rd_kafka_conf_t *conf;
char errstr[512];

conf = rd_kafka_conf_new();
rd_kafka_conf_set(conf, "compression.codec", "snappy", errstr, sizeof(errstr));
rd_kafka_conf_set(conf, "batch.num.messages", "100", errstr, sizeof(errstr));

rd_kafka_new(RD_KAFKA_PRODUCER, conf);

Threads and callbacks

librdkafka uses multiple threads internally to fully utilize modern hardware. The API is completely thread-safe and the calling application may call any of the API functions from any of its own threads at any time.

A poll-based API is used to provide signaling back to the application, the application should call rd_kafka_poll() at regular intervals. The poll API will call the following configured callbacks (optional):

  • dr_msg_cb - Message delivery report callback - signals that a message has been delivered or failed delivery, allowing the application to take action and to release any application resources used in the message.
  • error_cb - Error callback - signals an error. These errors are usually of an informational nature, i.e., failure to connect to a broker, and the application usually does not need to take any action. The type of error is passed as a rd_kafka_resp_err_t enum value, including both remote broker errors as well as local failures. An application typically does not have to perform any action when an error is raised through the error callback, the client will automatically try to recover from all errors, given that the client and cluster is correctly configured. In some specific cases a fatal error may occur which will render the client more or less inoperable for further use: if the error code in the error callback is set to RD_KAFKA_RESP_ERR__FATAL the application should retrieve the underlying fatal error and reason using the rd_kafka_fatal_error() call, and then begin terminating the instance. The Event API's EVENT_ERROR has a rd_kafka_event_error_is_fatal() function, and the C++ EventCb has a fatal() method, to help the application determine if an error is fatal or not.
  • stats_cb - Statistics callback - triggered if statistics.interval.ms is configured to a non-zero value, emitting metrics and internal state in JSON format, see [STATISTICS.md].
  • throttle_cb - Throttle callback - triggered whenever a broker has throttled (delayed) a request.

These callbacks will also be triggered by rd_kafka_flush(), rd_kafka_consumer_poll(), and any other functions that serve queues.

Optional callbacks not triggered by poll, these may be called spontaneously from any thread at any time:

  • log_cb - Logging callback - allows the application to output log messages generated by librdkafka.
  • partitioner - Partitioner callback - application provided message partitioner. The partitioner may be called in any thread at any time, it may be called multiple times for the same key. Partitioner function contraints:
    • MUST NOT call any rd_kafka_*() functions
    • MUST NOT block or execute for prolonged periods of time.
    • MUST return a value between 0 and partition_cnt-1, or the special RD_KAFKA_PARTITION_UA value if partitioning could not be performed.

Brokers

On initialization, librdkafka only needs a partial list of brokers (at least one), called the bootstrap brokers. The client will connect to the bootstrap brokers, specified by the bootstrap.servers (or metadata.broker.list) configuration property or by rd_kafka_brokers_add(), and query cluster Metadata information which contains the full list of brokers, topic, partitions and their leaders in the Kafka cluster.

Broker names are specified as host[:port] where the port is optional (default 9092) and the host is either a resolvable hostname or an IPv4 or IPv6 address. If host resolves to multiple addresses librdkafka will round-robin the addresses for each connection attempt. A DNS record containing all broker address can thus be used to provide a reliable bootstrap broker.

Sparse connections

Maintaining an open connection to each broker in the cluster is problematic for clusters with a large number of brokers, as well as for clusters with large number of clients.

To alleviate this, the enable.sparse.connections=true configuration property can be set to true (default), in which case the client will only connect to brokers it needs to communicate with, and only when necessary.

Examples of needed broker connections are:

  • leaders for partitions being consumed from
  • leaders for partitions being produced to
  • consumer group coordinator broker
  • cluster controller for Admin API operations
Random broker selection

When there is no broker connection and a connection to any broker is needed, such as on startup to retrieve metadata, the client randomly selects a broker from its list of brokers, which includes both the configure bootstrap brokers (including brokers manually added with rd_kafka_brokers_add()), as well as the brokers discovered from cluster metadata. Brokers with no prior connection attempt are tried first.

If there is already an available broker connection to any broker it is used, rather than connecting to a new one.

The random broker selection and connection scheduling is triggered when:

  • bootstrap servers are configured (rd_kafka_new())
  • brokers are manually added (rd_kafka_brokers_add()).
  • a consumer group coordinator needs to be found.
  • acquiring a ProducerID for the Idempotent Producer.
  • cluster or topic metadata is being refreshed.

A single connection attempt will be performed, and the broker will return to an idle INIT state on failure to connect.

The random broker selection is rate-limited to: 10 < reconnect.backoff.ms/2 < 1000 milliseconds.

Note: The broker connection will be maintained until it is closed by the broker (idle connection reaper).

Persistent broker connections

While the random broker selection is useful for one-off queries, there is need for the client to maintain persistent connections to certain brokers:

  • Consumer: the group coordinator.
  • Consumer: partition leader for topics being fetched from.
  • Producer: partition leader for topics being produced to.

These dependencies are discovered and maintained automatically, marking matching brokers as persistent, which will make the client maintain connections to these brokers at all times, reconnecting as necessary.

Connection close

A broker connection may be closed by the broker, intermediary network gear, due to network errors, timeouts, etc. When a broker connection is closed, librdkafka will back off the next reconnect attempt (to the given broker) for reconnect.backoff.ms -25% to +50% jitter, this value is increased exponentially for each connect attempt until reconnect.backoff.max.ms is reached, at which time the value is reset to reconnect.backoff.ms.

The broker will disconnect clients that have not sent any protocol requests within connections.max.idle.ms (broker configuration propertion, defaults to 10 minutes), but there is no fool proof way for the client to know that it was a deliberate close by the broker and not an error. To avoid logging these deliberate idle disconnects as errors the client employs some logic to try to classify a disconnect as an idle disconnect if no requests have been sent in the last socket.timeout.ms or there are no outstanding, or queued, requests waiting to be sent. In this case the standard "Disconnect" error log is silenced (will only be seen with debug enabled).

log.connection.close=false may be used to silence all disconnect logs, but it is recommended to instead rely on the above heuristics.

Feature discovery

Apache Kafka broker version 0.10.0 added support for the ApiVersionRequest API which allows a client to query a broker for its range of supported API versions.

librdkafka supports this functionality and will query each broker on connect for this information (if api.version.request=true) and use it to enable or disable various protocol features, such as MessageVersion 1 (timestamps), KafkaConsumer, etc.

If the broker fails to respond to the ApiVersionRequest librdkafka will assume the broker is too old to support the API and fall back to an older broker version's API. These fallback versions are hardcoded in librdkafka and is controlled by the broker.version.fallback configuration property.

Producer API

After setting up the rd_kafka_t object with type RD_KAFKA_PRODUCER and one or more rd_kafka_topic_t objects librdkafka is ready for accepting messages to be produced and sent to brokers.

The rd_kafka_produce() function takes the following arguments:

  • rkt - the topic to produce to, previously created with rd_kafka_topic_new()

  • partition - partition to produce to. If this is set to RD_KAFKA_PARTITION_UA (UnAssigned) then the configured partitioner function will be used to select a target partition.

  • msgflags - 0, or one of:

    • RD_KAFKA_MSG_F_COPY - librdkafka will immediately make a copy of the payload. Use this when the payload is in non-persistent memory, such as the stack.
    • RD_KAFKA_MSG_F_FREE - let librdkafka free the payload using free(3) when it is done with it.

    These two flags are mutually exclusive and neither need to be set in which case the payload is neither copied nor freed by librdkafka.

    If RD_KAFKA_MSG_F_COPY flag is not set no data copying will be performed and librdkafka will hold on the payload pointer until the message has been delivered or fails. The delivery report callback will be called when librdkafka is done with the message to let the application regain ownership of the payload memory. The application must not free the payload in the delivery report callback if RD_KAFKA_MSG_F_FREE is set.

  • payload,len - the message payload

  • key,keylen - an optional message key which can be used for partitioning. It will be passed to the topic partitioner callback, if any, and will be attached to the message when sending to the broker.

  • msg_opaque - an optional application-provided per-message opaque pointer that will be provided in the message delivery callback to let the application reference a specific message.

rd_kafka_produce() is a non-blocking API, it will enqueue the message on an internal queue and return immediately. If the number of queued messages would exceed the queue.buffering.max.messages configuration property then rd_kafka_produce() returns -1 and sets errno to ENOBUFS and last_error to RD_KAFKA_RESP_ERR__QUEUE_FULL, thus providing a backpressure mechanism.

rd_kafka_producev() provides an alternative produce API that does not require a topic rkt object and also provides support for extended message fields, such as timestamp and headers.

Note: See examples/rdkafka_performance.c for a producer implementation.

Simple Consumer API (legacy)

NOTE: For the high-level KafkaConsumer interface see rd_kafka_subscribe (rdkafka.h) or KafkaConsumer (rdkafkacpp.h)

The consumer API is a bit more stateful than the producer API. After creating rd_kafka_t with type RD_KAFKA_CONSUMER and rd_kafka_topic_t instances the application must also start the consumer for a given partition by calling rd_kafka_consume_start().

rd_kafka_consume_start() arguments:

  • rkt - the topic to start consuming from, previously created with rd_kafka_topic_new().
  • partition - partition to consume from.
  • offset - message offset to start consuming from. This may either be an absolute message offset or one of the two special offsets: RD_KAFKA_OFFSET_BEGINNING to start consuming from the beginning of the partition's queue (oldest message), or RD_KAFKA_OFFSET_END to start consuming at the next message to be produced to the partition, or RD_KAFKA_OFFSET_STORED to use the offset store.

After a topic+partition consumer has been started librdkafka will attempt to keep queued.min.messages messages in the local queue by repeatedly fetching batches of messages from the broker. librdkafka will fetch all consumed partitions for which that broker is a leader, through a single request.

This local message queue is then served to the application through three different consume APIs:

  • rd_kafka_consume() - consumes a single message
  • rd_kafka_consume_batch() - consumes one or more messages
  • rd_kafka_consume_callback() - consumes all messages in the local queue and calls a callback function for each one.

These three APIs are listed above the ascending order of performance, rd_kafka_consume() being the slowest and rd_kafka_consume_callback() being the fastest. The different consume variants are provided to cater for different application needs.

A consumed message, as provided or returned by each of the consume functions, is represented by the rd_kafka_message_t type.

rd_kafka_message_t members:

  • err - Error signaling back to the application. If this field is non-zero the payload field should be considered an error message and err is an error code (rd_kafka_resp_err_t). If err is zero then the message is a proper fetched message and payload et.al contains message payload data.
  • rkt,partition - Topic and partition for this message or error.
  • payload,len - Message payload data or error message (err!=0).
  • key,key_len - Optional message key as specified by the producer
  • offset - Message offset

Both the payload and key memory, as well as the message as a whole, is owned by librdkafka and must not be used after an rd_kafka_message_destroy() call. librdkafka will share the same messageset receive buffer memory for all message payloads of that messageset to avoid excessive copying which means that if the application decides to hang on to a single rd_kafka_message_t it will hinder the backing memory to be released for all other messages from the same messageset.

When the application is done consuming messages from a topic+partition it should call rd_kafka_consume_stop() to stop the consumer. This will also purge any messages currently in the local queue.

Note: See examples/rdkafka_performance.c for a consumer implementation.

Offset management

Broker based offset management is available for broker version >= 0.9.0 in conjunction with using the high-level KafkaConsumer interface (see rdkafka.h or rdkafkacpp.h)

Offset management is also available through a local offset file store, where the offset is periodically written to a local file for each topic+partition according to the following topic configuration properties:

  • auto.commit.enable
  • auto.commit.interval.ms
  • offset.store.path
  • offset.store.sync.interval.ms

There is currently no support for offset management with ZooKeeper.

Consumer groups

Broker based consumer groups (requires Apache Kafka broker >=0.9) are supported, see KafkaConsumer in rdkafka.h or rdkafkacpp.h

Topics

Topic auto creation

Topic auto creation is supported by librdkafka. The broker needs to be configured with auto.create.topics.enable=true.

Metadata

< 0.9.3

Previous to the 0.9.3 release librdkafka's metadata handling was chatty and excessive, which usually isn't a problem in small to medium-sized clusters, but in large clusters with a large amount of librdkafka clients the metadata requests could hog broker CPU and bandwidth.

> 0.9.3

The remaining Metadata sections describe the current behaviour.

Note: "Known topics" in the following section means topics for locally created rd_kafka_topic_t objects.

Query reasons

There are four reasons to query metadata:

  • brokers - update/populate cluster broker list, so the client can find and connect to any new brokers added.

  • specific topic - find leader or partition count for specific topic

  • known topics - same, but for all locally known topics.

  • all topics - get topic names for consumer group wildcard subscription matching

The above list is sorted so that the sub-sequent entries contain the information above, e.g., 'known topics' contains enough information to also satisfy 'specific topic' and 'brokers'.

Caching strategy

The prevalent cache timeout is metadata.max.age.ms, any cached entry will remain authoritative for this long or until a relevant broker error is returned.

  • brokers - eternally cached, the broker list is additative.

  • topics - cached for metadata.max.age.ms

Fatal errors

The added guarantee of ordering and no duplicates also requires a way for the client to fail gracefully when these guarantees can't be satisfied. If an unresolvable error occurs a fatal error is triggered in one or more of the follow ways depending on what APIs the application is utilizing:

  • C: the error_cb is triggered with error code RD_KAFKA_RESP_ERR__FATAL, the application should call rd_kafka_fatal_error() to retrieve the underlying fatal error code and error string.
  • C: an RD_KAFKA_EVENT_ERROR event is triggered and rd_kafka_event_error_is_fatal() returns true: the fatal error code and string are available through rd_kafka_event_error(), and .._string().
  • C++: an EVENT_ERROR event is triggered and event.fatal() returns true: the fatal error code and string are available through event.err() and event.str().

An application may call rd_kafka_fatal_error() at any time to check if a fatal error has been raised.

If a fatal error has been raised, sub-sequent use of the following API calls will fail:

  • rd_kafka_produce()
  • rd_kafka_producev()
  • rd_kafka_produce_batch()

The underlying fatal error code will be returned, depending on the error reporting scheme for each of those APIs.

When a fatal error has occurred the application should call rd_kafka_flush() to wait for all outstanding and queued messages to drain before terminating the application. rd_kafka_purge(RD_KAFKA_PURGE_F_QUEUE) is automatically called by the client when a producer fatal error has occurred, messages in-flight are not purged automatically to allow waiting for the proper acknowledgement from the broker. The purged messages in queue will fail with error code set to RD_KAFKA_RESP_ERR__PURGE_QUEUE.

Appendix

Test details

Test1: Produce to two brokers, two partitions, required.acks=2, 100 byte messages

Each broker is leader for one of the two partitions. The random partitioner is used (default) and each broker and partition is assigned approximately 250000 messages each.

Command:

# examples/rdkafka_performance -P -t test2 -s 100 -c 500000 -m "_____________Test1:TwoBrokers:500kmsgs:100bytes" -S 1 -a 2
....
% 500000 messages and 50000000 bytes sent in 587ms: 851531 msgs/s and 85.15 Mb/s, 0 messages failed, no compression

Result:

Message transfer rate is approximately 850000 messages per second, 85 megabytes per second.

Test2: Produce to one broker, one partition, required.acks=0, 100 byte messages

Command:

# examples/rdkafka_performance -P -t test2 -s 100 -c 500000 -m "_____________Test2:OneBrokers:500kmsgs:100bytes" -S 1 -a 0 -p 1
....
% 500000 messages and 50000000 bytes sent in 698ms: 715994 msgs/s and 71.60 Mb/s, 0 messages failed, no compression

Result:

Message transfer rate is approximately 710000 messages per second, 71 megabytes per second.

Test3: Produce to two brokers, two partitions, required.acks=2, 100 byte messages, snappy compression

Command:

# examples/rdkafka_performance -P -t test2 -s 100 -c 500000 -m "_____________Test3:TwoBrokers:500kmsgs:100bytes:snappy" -S 1 -a 2 -z snappy
....
% 500000 messages and 50000000 bytes sent in 1672ms: 298915 msgs/s and 29.89 Mb/s, 0 messages failed, snappy compression

Result:

Message transfer rate is approximately 300000 messages per second, 30 megabytes per second.

Test4: Produce to two brokers, two partitions, required.acks=2, 100 byte messages, gzip compression

Command:

# examples/rdkafka_performance -P -t test2 -s 100 -c 500000 -m "_____________Test3:TwoBrokers:500kmsgs:100bytes:gzip" -S 1 -a 2 -z gzip
....
% 500000 messages and 50000000 bytes sent in 2111ms: 236812 msgs/s and 23.68 Mb/s, 0 messages failed, gzip compression

Result:

Message transfer rate is approximately 230000 messages per second, 23 megabytes per second.