Large Payload Service

⚠️ This repository has been open sourced in order to help further discussion in the Temporal community about a proposal for official Temporal large payloads support.

We welcome experimentation and feedback, but please do not use in production. We make no guarantees about backwards compatibility of codecs, the HTTP interface, or storage drivers. If and when the upstream Temporal project includes support for large payloads, this repository will not be maintained.

This repository contains a HTTP service and accompanying Temporal Payload Codec which allows Temporal clients to automatically persist large payloads outside of workflow histories.

Temporal limits payload size to 4MB. The Large Payload Codec allows Temporal clients to read and write payloads larger than this limit by transparently storing and retrieving them via an HTTP service backed by cloud storage. In essence, the Large Payload Service implements a content addressed storage (CAS) interface for Temporal payloads.

For more details on the motivation, see the talk "Temporal at Datadog" from Replay 2022: https://youtu.be/LxgkAoTSI8Q?t=1686

• Usage
• Architecture
  • Example Conversion
  • Sequence Diagram for Payload Encoding
  • Sequence Diagram for Payload Decoding
• API
• Development

Usage

This repository does not provide any prebuilt binaries or images. The recommended approach is to build your own binary and image.

To programmatically build the Large Payload Service, you need to instantiate the driver and then pass it to server.NewHttpHandler. For example, to create a Large Payload Service instance backed by an S3 bucket, you would do something along these lines:

package main

import (
    "context"
    "net/http"
    "os"

    "github.com/aws/aws-sdk-go-v2/config"
    "github.com/DataDog/temporal-large-payload-codec/server"
    "github.com/DataDog/temporal-large-payload-codec/server/storage/s3"
    ...
)

func main() {
    region, set := os.LookupEnv("AWS_SDK_REGION")
    if !set {
        log.Fatal("AWS_REGION environment variable not set")
    }
    bucket, set := os.LookupEnv("BUCKET")
    if !set {
        log.Fatal("BUCKET environment variable not set")
    }

    cfg, err := config.LoadDefaultConfig(context.Background(), config.WithRegion(region))
    if err != nil {
        log.Fatal(err)
    }

    driver := s3.New(&s3.Config{
      Config: cfg,
      Bucket: bucket,
    })

    mux := http.NewServeMux()
    mux.Handle("/", server.NewHttpHandler(driver)))

    if err := http.ListenAndServe(":8577", mux); err != nil {
      log.Fatal(err)
    }
}

On the Temporal side, you need to create the Large Payload Server PayloadCodec, wrap it in a CodecDataConverter and pass it to the Temporal client contructor (simplified, without error handling):

opts := client.Options{
...
}

lpsEndpoint, _ := os.LookupEnv("LARGE_PAYLOAD_SERVICE_URL");
lpc, _ := largepayloadcodec.New(largepayloadcodec.WithURL(lpsEndpoint))
opts.DataConverter = converter.NewCodecDataConverter(opts.DataConverter, lpc)

temporalClient, _ := router.NewClient(opts)

Architecture

Architecturally, large payloads are passed through the CodecDataConverter which in turn uses the large payload codec to en- and decode the payloads. If the payload size is larger than the configured maximum payload size (default 128KB), the codec will use the Large Payload API to PUT or GET the payload from the Large Payload Service.

flowchart LR
	subgraph client["Temporal Client/Worker"]
		sdk[Temporal SDK]
		codec[[Large Payload<br/>Codec]]
	end

	subgraph service["Large Payload Service"]
		lps[Large Payload<br/>API]
		storage[(Cloud Storage)]
	end

	sdk -- Encode --> codec -. PUT blob .-> lps -. Write blob .-> storage
	sdk -- Decode --> codec -. GET blob .-> lps -. Read blob .-> storage

Example Conversion

Given a large input payload where data is larger than 128KB:

{
  "metadata": {
    "encoding": "text/plain"
  },
  "data": "..."
}

The large payload codec will encode this payload as follows:

{
  "metadata": {
    "encoding": "json/plain",
    "temporal.io/remote-codec": "v2"
  },
  "data": {
    "metadata": {"encoding": "text/plain"},
    "size": 1234567,
    "digest": "sha256:deadbeef",
    "key": "/blobs/default/sha256:deadbeef"
  }
}

• metadata: Original payload's metadata + temporal.io/remote-codec metadata header to indicate the use of the remote codec
• size: Size in bytes of data field in original payload
• digest: Digest of data in original payload (for integrity checks)
• key: Key used by the codec retrieve the stored payload

Sequence Diagram for Payload Encoding

sequenceDiagram
	autonumber

	participant SDK as Temporal SDK
	participant Converter as Dataconverter
	participant Codec as Large Payload<br/>Codec
	participant LPS as Large Payload<br/>Service
	participant Storage as Cloud Storage

	Note over SDK, Storage: Encoding a Payload

	SDK->>+Converter: Serialize Value
	Converter->>-SDK: Return Payload
	SDK->>+Codec: Encode Payload
	Codec->>Codec: Compute payload<br/>length & checksum
	Codec->>+LPS: PUT /v2/blobs
	LPS->>Storage: Verify checksum<br/>& write blob
	LPS->>-Codec: Return blob key
	Codec->>-SDK: Return (new) Payload

Sequence Diagram for Payload Decoding

sequenceDiagram
	autonumber
	participant SDK as Temporal SDK
	participant Converter as Dataconverter
	participant Codec as Large Payload<br/>Codec
	participant LPS as Large Payload<br/>Service
	participant Storage as Cloud Storage

	Note over SDK, Storage: Decoding a Payload

	SDK->>+Codec: Decode Payload
	Codec->>+LPS: GET /v2/blobs
	LPS->>Storage: Read blob
	LPS->>-Codec: Stream blob<br/>bytes
	Codec->>Codec: Verify checksum
	Codec->>-SDK: Return (original) Payload
	SDK->>+Converter: Deserialize Payload
	Converter->>-SDK: Return Value

API

The Large Payload Service offers the following API:

• /v2/health/head: Health check endpoint using a HEAD request.

  Returns the HTTP response status code 200 if the service is running correctly. Otherwise, an error code is returned.

• /v2/blobs/put: Upload endpoint expecting a PUT request.

  Required headers:

  • Content-Type set to application/octet-stream.
  • Content-Length set to the length of payload the data in bytes.
  • X-Temporal-Metadata set to the base64 encoded JSON of the Temporal Metadata.

  Query parameters:

  • namespace The Temporal namespace the client using the codec is connected to.

    The namespace forms part of the key for retrieval of the payload.

  • digest Specifies the checksum over the payload data using the format sha256:<sha256_hex_encoded_value>.

  The returned key of the put request needs to be stored and used for later retrieval of the payload. It is up to the Large Payload Server and the backend driver how to arrange the data in the backing data store. The server will honor, however, the value of remote-codec/key-prefix in the Temporal Metadata passed via the X-Temporal-Metadata header. It will use the specified string as prefix in the storage path.

• /v2/blobs/get: Download endpoint expecting a GET request.

  Required headers:

  • Content-Type set to application/octet-stream.
  • X-Payload-Expected-Content-Length set to the expected size of the payload data in bytes.

  Query parameters:

  • key specifying the key for the payload to retrieve.

Development

Refer to CONTRIBUTING.md for instructions on how to build and test the Large Payload Service and for general contributing guidelines.