Configuring Semantic Retry Policies (MoonBit)

Golem provides a composable, per-environment retry policy system. Policies are evaluated against error context properties and can be defined in the application manifest, managed via CLI, or created/overridden at runtime from agent code using the SDK.

1. Define Retry Policies in the Application Manifest

Add retry policy definitions under retryPolicyDefaults in golem.yaml, scoped per environment:

retryPolicyDefaults:
  prod:
    http-transient:
      priority: 10
      predicate:
        and:
          - propEq: { property: "error-type", value: "transient" }
          - propEq: { property: "uri-scheme", value: "https" }
      policy:
        countBox:
          maxRetries: 5
          inner:
            jitter:
              factor: 0.15
              inner:
                clamp:
                  minDelay: "100ms"
                  maxDelay: "5s"
                  inner:
                    exponential:
                      baseDelay: "200ms"
                      factor: 2.0
 
    catch-all:
      priority: 0
      predicate: true
      policy:
        countBox:
          maxRetries: 3
          inner:
            exponential:
              baseDelay: "100ms"
              factor: 3.0

Policy Evaluation Order

When an error occurs, policies are evaluated in descending priority order. The first matching predicate’s policy is applied. If no user-defined policy matches, the built-in default policy (3 retries, exponential backoff, clamped to [100ms, 1s], 15% jitter) is used.

Base Policies

Combinators

Predicates

Predicates are boolean expressions evaluated against error context properties. Compose with and, or, not:

• true / false — always/never match
• propEq — property equals a value
• propIn — property is one of a set of values
• propGte / propLt — numeric comparisons
• and / or / not — logical composition

Available Properties

Every retry decision happens in a specific context (an outgoing HTTP request, an HTTP response, a worker-to-worker RPC call, a trap from inside the guest, etc.). Each context only populates a subset of the property vocabulary below — a policy keyed on a property that is not present in the current context is silently skipped for that decision (it cannot apply there by definition).

Common to every context:

• verb — operation verb (HTTP method, RDBMS verb, RPC verb, or "trap" in the trap context)
• noun-uri — the resource URI (https://..., worker://..., kv://..., blobstore://..., dns://..., wasm://<function> for traps, golem://api, …)
• uri-scheme, uri-host, uri-port, uri-path — decomposed from noun-uri

Context-specific properties:

Practical consequence. A status-code-keyed policy (predicate: status-code in [...]) only fires for HTTP responses. The trap path does not see status-code and silently skips that policy — it does not error out. Likewise, a trap-type-keyed policy only fires from the trap path. Design one policy per context (or use or/and to make a policy explicitly match multiple contexts) rather than expecting a single policy to apply everywhere.

error-type values

• transient — transient transport failure (e.g. WASI HTTP error code, transient RDBMS error)
• http-status — HTTP response with a status code that matched a status-code-keyed policy

Status-code retries (opt-in)

Outgoing HTTP responses now flow through the retry-policy machinery: when the response arrives, its status-code is exposed to predicates. A policy is only considered for status-code retries if its predicate (or the predicate inside a nested FilteredOn) explicitly references the status-code property. Catch-all policies — including the synthesized default and any user-defined “matches all” policy — are intentionally excluded so status-based retries remain strictly opt-in.

When a matching policy decides to retry, the rejected response is dropped, the request body is reconstructed from the oplog, and the request is re-sent.

Eligibility rules (mirror inline transport retry):

• live execution (not replay/snapshot/PersistNothing)
• request body and trailers are reconstructible
• the HTTP method is idempotent, or assume_idempotence was set on the outgoing request
• not inside an atomically(...) block — in v1 status retries are skipped inside atomic regions; the user-land throw still triggers atomic-region replay, which gives equivalent end-to-end behavior

Example status-code policy:

http-5xx-retry:
  priority: 20
  predicate:
    and:
      - propIn: { property: "status-code", values: [500, 502, 503, 504] }
      - propEq: { property: "uri-scheme", value: "https" }
  policy:
    countBox:
      maxRetries: 3
      inner:
        exponential:
          baseDelay: "200ms"
          factor: 2.0

2. SDK: Build and Apply Retry Policies at Runtime

Use the NamedPolicy and Policy types to construct and apply retry policies from agent code:

let policy =
  NamedPolicy::named(
    "http-transient",
    Policy::exponential(Duration::millis(200), 2.0)
      .clamp(Duration::millis(100), Duration::seconds(5))
      .with_jitter(0.15)
      .only_when(
        Predicate::eq(Props::error_type(), Value::text("transient")),
      )
      .max_retries(5),
  )
    .priority(10)
    .applies_when(Predicate::eq(Props::uri_scheme(), Value::text("https")))

Scoped Usage with with_named_policy!

Apply a policy for a block of code — the previous policy is restored when the block exits:

with_named_policy!(policy, fn() {
  // HTTP calls in this block use the custom retry policy
  make_http_request()
})

Policy Builder Methods

Build policies fluently from base policies:

// Exponential backoff clamped with jitter and max retries
Policy::exponential(Duration::millis(200), 2.0)
  .clamp(Duration::millis(100), Duration::seconds(5))
  .with_jitter(0.15)
  .max_retries(5)
 
// Periodic with time limit
Policy::periodic(Duration::seconds(1))
  .time_box(Duration::seconds(60))
 
// Immediate retries then fall back to exponential
Policy::immediate()
  .max_retries(3)
  .and_then(
    Policy::exponential(Duration::seconds(1), 2.0)
      .max_retries(5)
  )
 
// Never retry (fail immediately)
Policy::never()

Predicate Builder Methods

// Match transient host-level failures
Predicate::eq(Props::error_type(), Value::text("transient"))
 
// Match a property value
Predicate::eq(Props::uri_scheme(), Value::text("https"))
 
// Combine predicates
Predicate::and_([
  Predicate::eq(Props::error_type(), Value::text("transient")),
  Predicate::eq(Props::uri_scheme(), Value::text("https")),
])

3. Querying Retry Policies at Runtime

Use the query API to inspect active policies from agent code:

// List all active policies
let policies = @api.get_retry_policies()
policies.each(fn(p) {
  @log.info("Policy '\{p.name}' priority=\{p.priority}")
})
 
// Get a specific policy by name
match @api.get_retry_policy_by_name("http-transient") {
  Some(policy) => @log.info("Found policy with priority \{policy.priority}")
  None => @log.warn("Policy not found")
}

The returned NamedRetryPolicy has fields: name (String), priority (UInt), predicate (RetryPredicate), policy (RetryPolicy).

4. Live-Editing Policies via CLI

Retry policies can be managed at runtime without redeployment:

# Create a new policy
golem retry-policy create http-transient \
  --priority 10 \
  --predicate '{ "and": [{ "propEq": { "property": "error-type", "value": "transient" } }, { "propEq": { "property": "uri-scheme", "value": "https" } }] }' \
  --policy '{ "countBox": { "maxRetries": 5, "inner": { "exponential": { "baseDelay": "200ms", "factor": 2.0 } } } }'
 
# List all policies in the current environment
golem retry-policy list
 
# Get a specific policy by name
golem retry-policy get http-transient
 
# Update an existing policy
golem retry-policy update http-transient --priority 15
 
# Delete a policy
golem retry-policy delete http-transient

5. Default Retry Policy

When no user-defined retry policies are set, Golem activates a default catch-all:

• Name: default
• Priority: 0
• Predicate: true (matches everything)
• Policy: Up to 3 retries, exponential backoff (factor 3.0), delays clamped to [100ms, 1s], 15% jitter

Key Constraints

• Policies are defined per-environment — different environments can have different retry behaviors
• Policy names must be unique within an environment
• Higher priority policies are evaluated first; the first matching predicate wins
• with_named_policy! is scoped — the policy is restored when the closure exits
• Inline retries (automatic transparent retries for transient network errors) happen before the policy system kicks in
• Changes made via CLI or REST API take effect immediately for running agents