# ADR-001: Circuit Breaker and Retry via gobreaker and avast/retry-go

**Status:** Accepted
**Date:** 2026-03-18

## Context

Outbound HTTP calls to external services are subject to transient failures (network blips,
brief service restarts) and sustained failures (outages, overloads). Two complementary
strategies address these cases:

- **Retry** recovers from transient failures by re-attempting the request a limited number
  of times before giving up.
- **Circuit breaking** detects sustained failure patterns and stops sending requests to a
  failing service, giving it time to recover and preventing the caller from accumulating
  blocked goroutines.

Implementing both from scratch introduces risk of subtle bugs (backoff arithmetic, state
machine transitions). Well-tested, widely adopted libraries are preferable.

## Decision

Two external libraries are composed:

**Retry: `github.com/avast/retry-go/v4`**
- Configured via `Config.MaxRetries` and `Config.RetryDelay`.
- Uses `retry.BackOffDelay` (exponential backoff) to avoid hammering a failing service.
- `retry.LastErrorOnly(true)` ensures only the final error from the retry loop is reported.
- Only HTTP 5xx responses trigger a retry. 4xx responses are not retried (they represent
  caller errors, not server instability).

**Circuit breaker: `github.com/sony/gobreaker`**
- Configured via `Config.CBThreshold` (consecutive failures to trip) and `Config.CBTimeout`
  (time in open state before transitioning to half-open).
- The retry loop runs inside the circuit breaker's `Execute` call. A full retry sequence
  counts as one attempt from the circuit breaker's perspective only if all retries fail.
- When the circuit opens, `Do` returns `xerrors.ErrUnavailable` immediately, without
  attempting the network call.
- State changes are logged via the duck-typed `Logger` interface.

The nesting order (circuit breaker wraps retry) is intentional: the circuit breaker
accumulates failures at the level of "did the request ultimately succeed after retries",
not at the level of individual attempts.

## Consequences

**Positive:**
- Transient failures are handled transparently by the caller.
- Sustained outages are detected quickly and the circuit opens, returning fast errors.
- Configuration is explicit and environment-variable driven.
- Circuit state changes are observable via logs.

**Negative:**
- Retry with backoff increases total latency for failing requests up to
  `MaxRetries * RetryDelay * (2^MaxRetries - 1)` in the worst case.
- The circuit breaker counts only consecutive failures (`ConsecutiveFailures >= CBThreshold`),
  not a rolling failure rate. Interleaved successes reset the counter.
- `gobreaker.ErrOpenState` is wrapped in `xerrors.ErrUnavailable`, so callers must check for
  this specific code to distinguish circuit-open from normal 503 responses.