Presto Worker REST API#

Presto’s coordinator communicates with Presto workers to execute query fragments and fetch query results. Presto workers communicate with each other to exchange intermediate results. This chapter documents the REST API used in these communications.

Task resource is used to start execution of a query fragment, track status, and fetch results.

Control Plane#

The following HTTP methods are used by the coordinator to start execution of a query fragment and track execution status.

  • A POST to /v1/task/{taskId} starts execution of the query fragment specified in the POST body. The request optionally includes a set of initial splits to process. The request also specifies how to partition results, e.g. hash partition using specified output columns into specified number of output buffers or combine all results into a single output buffer or broadcast combined results into multiple output buffers.

  • A subsequent POST to /v1/task/{taskId} may provide additional splits for processing and eventually specify that no more splits will be coming.

  • A GET on /v1/task/{taskId}/status returns a TaskStatus JSON document describing the current execution status.

  • A GET on /v1/task/{taskId} returns a TaskInfo JSON document containing extended information about the execution status.

  • A DELETE on /v1/task/{taskId} deletes a finished task or cancels a task in-progress.

  • A GET on /v1/task returns JSON document containing the list of TaskInfo for all tasks.

A status request from the coordinator includes two HTTP headers: X-Presto-Current-State and X-Presto-Max-Wait. X-Presto-Current-State specifies the task state known to the coordinator. If the task state on the worker is different, the worker will reply immediately. IF the task state on the worker is the same as on coordinator, the worker will wait for the task state to change before replying. X-Presto-Max-Wait HTTP header specifies the maximum wait time. The worker will reply after that much time even if the task state stays the same.

The design ensures that coordinator receives task state changes in a timely manner without polling the worker in a tight loop.

The same design applies to requests for extended task information via a GET on /v1/task/{taskId}.

Data Plane#

The following HTTP methods are used by the coordinator to fetch final query results or by a downstream worker to fetch intermediate results from the upstream worker.

  • A GET on {taskId}/results/{bufferId}/{token} returns the next batch of results from the specified output buffer.

  • A GET on {taskId}/results/{bufferId}/{token}/acknowledge acknowledges the receipt of the results and allows the worker to delete them.

  • A DELETE on {taskId}/results/{bufferId} deletes all results from the specified output buffer in case of an error.

Coordinator and workers fetch results in chunks. They specify the maximum size in bytes for the chunk using X-Presto-Max-Size HTTP header. Each chunk is identified by a monotonically increasing sequence number sometimes referred to as a token. The first request for results specifies sequence number zero. The response includes:

  • The requested sequence number as X-Presto-Page-Sequence-Id HTTP header,

  • The sequence number to use to acknowledge the receipt of the chunk and to request the next chunk as X-Presto-Page-End-Sequence-Id HTTP header,

  • An indication that there are no more results as X-Presto-Buffer-Complete HTTP header with the value of “true”.

The body of the response contains a list of pages in SerializedPage wire format.

After receiving the first chunk of results, the client sends an ack via a GET on {taskId}/results/{bufferId}/{token}/acknowledge with the token set to the value of the X-Presto-Page-End-Sequence-Id HTTP header received earlier along with the results. Then, the client uses that sequence number to request the next chunk of results. The client keeps fetching results until it receives X-Presto-Buffer-Complete HTTP header with the value of “true”.

If the worker times out populating a response, or the task has already failed or been aborted, the worker will return empty results. The client can attempt to retry the request. In the case where the task is in a terminal state, it is assumed that the Control Plane will eventually handle the state change.

If the client missed a response it can repeat the request and the worker will send the results again. Upon receiving an ack for a sequence number, the worker deletes all results with the sequence number less than that and the client can no longer re-fetch these.

Here is a sample message passing diagram for fetching two-chunk result from the output buffer zero.

../_images/worker-protocol-results.png

Output Buffers#

Data shuffle involves workers in a downstream stage fetching results from the workers in an upstream stage. Each producing upstream worker sets up as many output buffers as there are workers in the downstream stage. Output buffers are identified by a sequential numbers starting with zero. Each downstream worker is assigned a single output buffer and it uses it to fetch results from all the upstream workers.

The following diagram shows 3 downstream workers. These are assigned output buffer numbers 0, 1, and 2. Each upstream worker has 3 output buffers. The downstream worker #0 fetches results from all the upstream workers using buffer number 0. The downstream worker #1 fetches results from all the upstream workers using buffer number 1. The downstream worker #2 fetches results from all the upstream workers using buffer number 2.

../_images/worker-protocol-output-buffers.png