tasks/docs/architecture/retry.md

Retry

The go.temporal.io/server/common/backoff package contains the retry primitives used across Temporal.

backoff.ThrottleRetry or, when there is a context.Context available backoff.ThrottleRetryContext, can be used to retry any operation that returns an error.

Both are configured via:

• backoff.IsRetryable which decides whether to retry based on the error type
• backoff.RetryPolicy which decides how long to backoff first - or not retry at all

It's important to note that a special retry policy is used for ResourceExhausted service errors.

Service Error

Service errors are specific Go errors that can generate a gRCP Status (see status.proto). A gRPC status contains a gRPC Code (see code.proto), a message and (optionally) a payload with more details.

type ServiceError interface {
    error
    Status() *status.Status
}

The api-go repository defines most service errors:

• general-purpose errors (such as Canceled, NotFound or Unavailable)
• specialized errors which carry more details (such as NamespaceNotActive with the gRPC code FailedPrecondition)

Furthermore, a few more Server-specific service errors are defined in this repository, such as ShardOwnershipLost or TaskAlreadyStarted.

gRPC

A failed gRPC request can be retried by the (server-side) gRPC handler and the service client. Both use the aforementioned backoff package to configure and execute the retries.

gRPC handler: All Temporal services wrap their API handlers with the gRPC interceptor interceptor.RetryableInterceptor to retry a failed gRPC request with certain errors. Look for NewRetryableInterceptor to see the configuration for each service.

gRPC client: Similarly, each service client can retry a failed gRPC request with certain errors. Look for NewRetryableClient to see the configuration for each service client.

For example, for a gRPC request from Frontend to History, there are 3 places an error can be retried (marked with 🔁):

NOTE: gRPC handler retries can be more efficient since they avoid a round-trip, but note that retry behavior multiplies between server and client, and client-side is more flexible (e.g. can direct a request to a different server). So gRPC handler retries should be used sparingly and are configured to do no more than one extra attempt.