alpaca-broker-rate-limits-resilience

Alpaca — Rate Limits & Resilience

Alpaca's APIs are rate-limited and occasionally flaky under load. Any client that does more than a handful of calls — especially bulk jobs, backfills, and reconciliation sweeps — needs disciplined retry, backoff, and concurrency control. These patterns are transport-level and apply in any language.

Read alpaca-broker-integration first.

1. Rate-limit headers — read them on every response

Alpaca returns standard headers:

Header	Meaning
X-RateLimit-Limit	requests allowed in the window
X-RateLimit-Remaining	requests left in the current window
X-RateLimit-Reset	unix timestamp (seconds) when the window resets

Parse them on every response, not just on errors. Two uses:

• Proactive: when Remaining drops below a threshold (e.g. ≤ 50), log a warning and/or slow down — you're about to get throttled.
• Reactive: on 429, use Reset to wait exactly until the window opens.

Limits vary by endpoint and plan; market-data limits differ from broker limits. Don't hardcode a number — react to the headers.

2. The retry loop (pseudocode)

MAX_ATTEMPTS = 10
INITIAL_DELAY_MS = 1000

for attempt in 1..MAX_ATTEMPTS:
    res = http(request)                      # with a sane timeout (see §5)
    remaining, reset_at = parse_rate_headers(res.headers)
    if remaining <= 50: log_warn("approaching rate limit", reset_at)

    if res.status == 429:
        # wait until the window resets, plus a small buffer
        wait = (reset_at - now()) if reset_at else INITIAL_DELAY_MS * 2^(attempt-1)
        sleep(max(0, wait) + 1000)           # +1s buffer past reset
        continue

    if res.status in (500, 502, 503, 504) or network_error:
        sleep(INITIAL_DELAY_MS * 2^(attempt-1))   # exponential backoff
        continue

    return res                                # success or non-retryable 4xx
raise last_error

Key points:

• On 429, wait until X-RateLimit-Reset + a ~1s buffer — don't blindly exponential-backoff when the API told you exactly when to retry.
• Exponential backoff (base * 2^(attempt-1)) for network errors and 5xx. With base 1s and 10 attempts the tail is minutes — fine for background jobs, too slow for user-facing calls (use fewer attempts there).
• Don't retry non-retryable 4xx (400/403/422) — those won't fix themselves; surface them.
• Optionally add jitter to backoff to avoid thundering-herd when many workers retry together.