Pulsar binary protocol specification

Pulsar uses a custom binary protocol for communications between producers/consumers and brokers. The protocol is designed to support all the required features (eg: acknowledgments, flow control) while ensuring maximum transport and implementation efficiency.

Clients and brokers exchange commands between each other. Commands are formatted as binary Protocol Buffers messages.

The format of protobuf commands is specified in the PulsarApi.proto file pulsar-common/src/main/proto/PulsarApi.proto

Commands for different producers and consumers can be interleaved and sent through the same connection without any restriction.

All the commands are embedded in a BaseCommand protobuf object that includes a type enum and all the possible sub-commands as optional fields. At any time, one BaseCommand can only have set a single sub-command.

Framing

Since protobuf doesn't provide any sort of message frame, we add a 4 bytes size field before the protobuf data.

The maximum size for a single frame is 5 Mbytes.

In Pulsar protocol we have 2 types of commands:

  1. Simple commands that do not carry payloads
  2. Commands with payload, used when publishing or delivering messages. In this case the protobuf command data is followed by some other protobuf metadata and then the payload which is passed in raw format, outside of protobuf for efficiency reasons.

All sizes are passed as 4 bytes unsigned big endian integers.

Simple commands

[TOTAL_SIZE] [CMD_SIZE] [CMD]
  • TOTAL_SIZE → (4 bytes) Size of the frame in bytes. Counting everything that comes after it
  • CMD_SIZE → (4 bytes) Size of the protobuf serialized command
  • CMD → Protobuf object serialized in binary form

Payload commands

[TOTAL_SIZE] [CMD_SIZE][CMD] [MAGIC_NUMBER][CHECKSUM] [METADATA_SIZE][METADATA] [PAYLOAD]
  • TOTAL_SIZE → (4 bytes) Size of the frame in bytes. Counting everything that comes after it
  • CMD_SIZE → (4 bytes) Size of the protobuf serialized command
  • CMD → Protobuf object serialized in binary form
  • Anything coming after it's what gets actually persisted
    • MAGIC_NUMBER → (2 bytes) 0x0e01 Magic number identifying the current format
    • CHECKSUM → (2 bytes) CRC32-C checksum of anything that comes after it
    • METADATA_SIZE → (4 bytes) Size of the message metadata
    • METADATA → Message metadata stored as binary protobuf message
    • PAYLOAD → Anything left in the frame is considered to be the payload and it can include any sequence of bytes

Message metadata

Message metadata is stored alongside with the application specified payload as a serialized protobuf object.

Metadata is created by the producer, and passed on unchanged to the consumer.

Fields:

  • producer_name → Name of the producer that published the message
  • sequence_id → Sequence id of the message, assigned by producer
  • publish_time → Publish timestamp. Number of ms since Jan 1st 1970 in UTC
  • properties → Sequence of Pair<String, String>. These are application defined keys and values with no meaning to Pulsar
  • replicated_from → (optional) Indicated that the message has been replicated and specifies the name of the cluster where the message was originally published
  • partition_key → (optional) While publishing on a partition topic, if the key is present, the hash of the key is used to determine which partition to choose
  • compression → (optional) Signal that payload has been compressed and with which compression library
  • uncompressed_size → (optional) If compression is used, producer must fill the uncompressed size field with the original payload size
  • num_messages_in_batch → (optional) If this message is really a batch of multiple entries, this field must be set to the number of messages in the batch

Batch messages

When using batch messages, the payload will be containing a list of entries, each of them with its individual metadata, defined by the SingleMessageMetadata object.

For a single batch, the payload format will look like:

[MD_SIZE_1] [MD_1] [PAYLOAD_1] [MD_SIZE_2] [MD_2] [PAYLOAD_2] ...

With:

  • MD_SIZE_X → Size of the single message metadata serialized protobuf
  • MD_X → Single message metadata
  • PAYLOAD_X → Message payload passed by application

Fields for SingleMessageMetadata:

  • properties → Application defined properties
  • partition_key → (optional) Key to indicate the hashing to a particular partition
  • payload_size → Size of the payload for the single message in the batch

When compression is enabled, the whole batch will be compressed at once.

Interactions

Connection establishment

After opening a TCP connection to a broker, typically on port 6650, the client is responsible to initiate the session.

![Connect interaction](img/Binary Protocol - Connect.png)

After receiving a Connected response from the broker, the client can consider the connection ready to use. Alternatively, if the broker doesn't validate the client authentication, it will reply with an Error command and close the TCP connection.

Example:

CommandConnect {
   "client_version" : "Pulsar-Client-Java-v1.15.2",
   "auth_method_name" : "my-authentication-plugin",
   "auth_data" : "my-auth-data",
   "protocol_version" : 6
}

Fields:

  • client_version → String based identifier. Format is not enforced
  • auth_method_name → (optional) Name of the authentication plugin if auth enabled
  • auth_data → (optional) Plugin specific authentication data
  • protocol_version → Indicates the protocol version supported by the client. Broker will not send commands introduced in newer revisions of the protocol. Broker might be enforcing a minimum version
CommandConnected {
   "server_version" : "Pulsar-Broker-v1.15.2",
   "protocol_version" : 6
}

Fields:

  • server_version → String identifier of broker version
  • protocol_version → Protocol version supported by the broker. Client must not attempt to send commands introduced in newer revisions of the protocol

Keep Alive

To identify prolonged network partitions between clients and brokers or cases in which a machine crashes without interrupting the TCP connection on the remote end (eg: power outage, kernel panic, hard reboot...), we have introduced a mechanism to probe for the availability status of the remote peer.

Both clients and brokers are sending Ping commands periodically and they will close the socket if a Pong response is not received within a timeout (default used by broker is 60s).

A valid implementation of a Pulsar client is not required to send the Ping probe, though it is required to promptly reply after receiving one from the broker, in order to prevent the remote side to forcibly close the TCP connection.

Producer

In order to send messages, a client needs to establish a producer. When creating a producer, the broker will first verify that this particular client is authorized to publish on the topic.

Once the client gets confirmation of the producer creation, it can publish messages to the broker, referring to the producer id negotiated before.

![Producer interaction](img/Binary Protocol - Producer.png)

Command Producer
CommandProducer {
  "topic" : "persistent://my-property/my-cluster/my-namespace/my-topic",
  "producer_id" : 1,
  "request_id" : 1
}

Parameters:

  • topic → Complete topic name to where you want to create the producer on
  • producer_id → Client generated producer identifier. Needs to be unique within the same connection
  • request_id → Identifier for this request. Used to match the response with the originating request. Needs to be unique within the same connection
  • producer_name → (optional) If a producer name is specified, the name will be used, otherwise the broker will generate a unique name. Generated producer name is guaranteed to be globally unique. Implementations are expected to let the broker generate a new producer name when the producer is initially created, then reuse it when recreating the producer after reconnections.

The broker will reply with either ProducerSuccess or Error commands.

Command ProducerSuccess
CommandProducerSuccess {
	"request_id" :  1,
	"producer_name" : "generated-unique-producer-name"
}

Parameters:

  • request_id → Original id of the CreateProducer request
  • producer_name → Generated globally unique producer name or the name specified by the client, if any.
Command Send

Command Send is used to publish a new message within the context of an already existing producer. This command is used in a frame that includes command as well as message payload, for which the complete format is specified in the payload commands section.

CommandSend {
	"producer_id" : 1,
	"sequence_id" : 0,
	"num_messages" : 1
}

Parameters:

  • producer_id → id of an existing producer
  • sequence_id → each message has an associated sequence id which is expected to be implemented with a counter starting at 0. The SendReceipt that acknowledges the effective publishing of a messages will refer to it by its sequence id.
  • num_messages → (optional) Used when publishing a batch of messages at once.
Command SendReceipt

After a message has been persisted on the configured number of replicas, the broker will send the acknowledgment receipt to the producer.

CommandSendReceipt {
	"producer_id" : 1,
	"sequence_id" : 0,
	"message_id" : {
     "ledgerId" : 123,
     "entryId" : 456
  }
}

Parameters:

  • producer_id → id of producer originating the send request
  • sequence_id → sequence id of the published message
  • message_id → message id assigned by the system to the published message Unique within a single cluster. Message id is composed of 2 longs, ledgerId and entryId, that reflect that this unique id is assigned when appending to a BookKeeper ledger
Command CloseProducer

Note: This command can be sent by either producer or broker.

When receiving a CloseProducer command, the broker will stop accepting any more messages for the producer, wait until all pending messages are persisted and then reply Success to the client.

The broker can send a CloseProducer command to client when it's performing a graceful failover (eg: broker is being restarted, or the topic is being unloaded by load balancer to be transferred to a different broker).

When receiving the CloseProducer, the client is expected to go through the service discovery lookup again and recreate the producer again. The TCP connection is not being affected.

Consumer

A consumer is used to attach to a subscription and consume messages from it. After every reconnection, a client needs to subscribe to the topic. If a subscription is not already there, a new one will be created.

![Consumer](img/Binary Protocol - Consumer.png)

Flow control

After the consumer is ready, the client needs to give permission to the broker to push messages. This is done with the Flow command.

A Flow command gives additional permits to send messages to the consumer. A typical consumer implementation will use a queue to accumulate these messages before the application is ready to consume them.

After the application has dequeued a number of message, the consumer will send additional number of permits to allow the broker to push more messages.

Command Subscribe
CommandSubscribe {
	"topic" : "persistent://my-property/my-cluster/my-namespace/my-topic",
	"subscription" : "my-subscription-name",
	"subType" : "Exclusive",
	"consumer_id" : 1,
	"request_id" : 1
}

Parameters:

  • topic → Complete topic name to where you want to create the consumer on
  • subscription → Subscription name
  • subType → Subscription type: Exclusive, Shared, Failover
  • consumer_id → Client generated consumer identifier. Needs to be unique within the same connection
  • request_id → Identifier for this request. Used to match the response with the originating request. Needs to be unique within the same connection
  • consumer_name → (optional) Clients can specify a consumer name. This name can be used to track a particular consumer in the stats. Also, in Failover subscription type, the name is used to decide which consumer is elected as master (the one receiving messages): consumers are sorted by their consumer name and the first one is elected master.
Command Flow
CommandFlow {
  "consumer_id" : 1,
  "messagePermits" : 1000
}

Parameters:

  • consumer_id → Id of an already established consumer
  • messagePermits → Number of additional permits to grant to the broker for pushing more messages
Command Message

Command Message is used by the broker to push messages to an existing consumer, within the limits of the given permits.

This command is used in a frame that includes the message payload as well, for which the complete format is specified in the payload commands section.

CommandMessage {
	"consumer_id" : 1,
	"message_id" : {
    "ledgerId" : 123,
    "entryId" : 456
  }
}

Command Ack

An Ack is used to signal to the broker that a given message has been successfully processed by the application and can be discarded by the broker.

In addition, the broker will also maintain the consumer position based on the acknowledged messages.

message CommandAck {
	"consumer_id" : 1,
	"ack_type" : "Individual",
	"message_id" : {
    "ledgerId" : 123,
    "entryId" : 456
  }
}

Parameters:

  • consumer_id → Id of an already established consumer
  • ack_type → Type of acknowledgment: Individual or Cumulative
  • message_id → Id of the message to acknowledge
  • validation_error → (optional) Indicates that the consumer has discarded the messages due to: UncompressedSizeCorruption, DecompressionError, ChecksumMismatch, BatchDeSerializeError
Command CloseConsumer

Note: This command can be sent by either producer or broker.

This command behaves the same as CloseProducer

Command RedeliverUnacknowledgedMessages

A consumer can ask the broker to redeliver some or all of the pending messages that were pushed to that particular consumer and not yet acknowledged.

The protobuf object accepts a list of message ids that the consumer wants to be redelivered. If the list is empty, the broker will redeliver all the pending messages.

On redelivery, messages an be sent to the same consumer or, in the case of a shared subscription, spread across all available consumers.

Service discovery

Topic lookup

Topic lookup needs to be performed each time a client needs to create or reconnect a producer or a consumer. Lookup is used to discover which particular broker is serving the topic we are about to use.

Lookup can be done with a REST call as described in the admin API docs.

Since Pulsar-1.16 it is also possible to perform the lookup within the binary protocol.

For the sake of example, let's assume we have a service discovery component running at pulsar://broker.example.com:6650

Individual brokers will be running at pulsar://broker-1.example.com:6650, pulsar://broker-2.example.com:6650, ...

A client can use a connection to the discovery service host to issue a LookupTopic command. The response can either be a broker hostname to connect to, or a broker hostname to which retry the lookup.

The LookupTopic command has to be used in a connection that has already gone through the Connect / Connected initial handshake.

![Topic lookup](img/Binary Protocol - Topic lookup.png)

CommandLookupTopic {
	"topic" : "persistent://my-property/my-cluster/my-namespace/my-topic",
	"request_id" : 1,
	"authoritative" : false
}

Fields:

  • topic → Topic name to lookup
  • request_id → Id of the request that will be passed with its response
  • authoritative → Initial lookup request should use false. When following a redirect response, client should pass the same value contained in the response
LookupTopicResponse

Example of response with successful lookup:

Command LookupTopicResponse {
  "request_id" : 1,
  "response" : "Connect",
  "brokerServiceUrl" : "pulsar://broker-1.example.com:6650",
  "brokerServiceUrlTls" : "pulsar+ssl://broker-1.example.com:6651",
  "authoritative" : true
}

Example of lookup response with redirection:

Command LookupTopicResponse {
  "request_id" : 1,
  "response" : "Redirect",
  "brokerServiceUrl" : "pulsar://broker-2.example.com:6650",
  "brokerServiceUrlTls" : "pulsar+ssl://broker-2.example.com:6651",
  "authoritative" : true
}

In this second case, we need to reissue the LookupTopic command request to broker-2.example.com and this broker will be able to give a definitive answer to the lookup request.

Partitioned topics discovery

Partitioned topics metadata discovery is used to find out if a topic is a “partitioned topic” and how many partitions were set up.

If the topic is marked as “partitioned”, the client is expected to create multiple producers or consumers, one for each partition, using the partition-X suffix.

This information only needs to be retrieved the first time a producer or consumer is created. There is no need to do this after reconnections.

The discovery of partitioned topics metadata works very similar to the topic lookup. The client send a request to the service discovery address and the response will contain actual metadata.

Command PartitionedTopicMetadata
CommandPartitionedTopicMetadata {
  "topic" : "persistent://my-property/my-cluster/my-namespace/my-topic",
  "request_id" : 1
}

Fields:

  • topic → the topic for which to check the partitions metadata
  • request_id → Id of the request that will be passed with its response
Command PartitionedTopicMetadataResponse

Example of response with metadata:

CommandPartitionedTopicMetadataResponse {
  "request_id" : 1,
  "response" : "Success",
  "partitions" : 32
}