Api Connector Builder

API connector builder — auth, pagination, retries, rate limits

You are writing a client for someone else's HTTP API. You do not own the contract; you obey it. The deliverable is one typed connector module per vendor that authenticates, walks the whole result set, retries only transient failures with backoff, and stays under the rate limit without getting the key banned. One connector = one vendor: mixed clients tangle two auth schemes and two rate-limit budgets into something no one can reason about.

Four pillars, every time: auth, pagination, retries, rate limits. If your connector skips any one of them it works in the demo and breaks in production — on page 2, on token expiry, on the first 429, or on a flaky network.

Step 0 — read the contract

Read the vendor docs before writing a line — invented endpoints and guessed field names 404 in prod. Extract these first; each one changes what you write, so missing one means a rewrite.

If a fact is not in the docs, probe one real call and read the response headers and body — do not assume.

Auth — pick the flow, then store the secret right

OAuth 2.1 is the current baseline. Three deltas you must honor: PKCE is mandatory for every authorization-code client, the Implicit and Resource-Owner-Password grants are removed, and bearer tokens may not travel in query strings (header only) — query strings end up in logs and referrers. (oauth.net/2.1, accessed 2026-06-02.)

Refresh-token rule (public clients): refresh tokens must be sender- constrained or one-time-use — rotated on every refresh, with the old one invalidated. Keep access tokens short-lived. Never log a raw token; log a partial or hash if you must correlate. (OAuth 2.0 Security BCP / RFC 9700, accessed 2026-06-02.) Full per-flow walkthroughs, token storage, and a DPoP note are in references/auth-flows.md.

Retries + backoff

Retry only transient failures, and only when the operation is safe to repeat.

Idempotency gate. GET/PUT/DELETE are idempotent by semantics and safe to retry. POST is not — only retry it if you send a stable idempotency key so the server dedupes the duplicate. (AWS Builders' Library, accessed 2026-06-02.)

Backoff with jitter. Use delay = min(cap, base * 2^attempt) + random_jitter. The jitter is not optional: without it, every client that failed at the same instant retries at the same instant — a synchronized thundering herd that DDoSes the recovering server. Full or decorrelated jitter is preferred. Always set stop conditions: max attempts 3–5, a total deadline, and a per-attempt timeout. (AWS Builders' Library, accessed 2026-06-02.)

pythonCopy code
# Bad: retries everything, flat sleep, no jitter, no cap, no deadline.
for _ in range(10):
    r = httpx.get(url)
    if r.status_code == 200:
        return r.json()
    time.sleep(1)  # 4xx will never recover; herds synchronize on flat 1s

pythonCopy code
# Good: transient-only, exponential backoff WITH jitter, bounded.
import random, time, httpx

RETRYABLE = {429, 502, 503, 504}

def get(url, *, attempts=5, base=0.5, cap=20.0, deadline=60.0):
    started = time.monotonic()
    for attempt in range(attempts):
        try:
            r = httpx.get(url, timeout=10.0)  # per-attempt timeout
        except (httpx.ConnectError, httpx.ReadTimeout):
            pass  # network transient -> fall through to backoff
        else:
            if r.status_code == 200:
                return r.json()
            if r.status_code not in RETRYABLE:
                r.raise_for_status()  # 4xx: do not retry, surface it
        if time.monotonic() - started > deadline:
            raise TimeoutError("retry deadline exceeded")
        delay = min(cap, base * 2 ** attempt) + random.uniform(0, base)
        time.sleep(delay)
    raise RuntimeError("max attempts exhausted")

In Python prefer tenacity 9.1.4 (@retry with wait_exponential_jitter, stop_after_attempt, retry_if_exception_type) over a hand loop; in Node/TS use undici (the engine behind global fetch() since Node 18) with its retry interceptor. (PyPI/tenacity, nodejs/undici, accessed 2026-06-02.)

Rate limits

Do not guess the wait — the server tells you. Precedence:

1. Retry-After present (on a 429 or 503) → wait exactly that. It is either a number of seconds or an HTTP-date; handle both.
2. No Retry-After → compute the wait from X-RateLimit-Reset (a Unix epoch or a delta-seconds, per vendor docs).
3. Proactively throttle on X-RateLimit-Remaining — slow down before you hit zero rather than absorbing a wall of 429s. (iotools.cloud rate-limiting guidance, accessed 2026-06-02.)

pythonCopy code
# Bad: ignore the headers, hammer, eat 429s, get the key throttled or banned.
while more:
    resp = client.get(next_url)   # no remaining check, no Retry-After
    process(resp)

pythonCopy code
# Good: header-aware. Honor Retry-After, else reset; brake near the limit.
def wait_for_rate_limit(resp):
    ra = resp.headers.get("Retry-After")
    if ra is not None:
        return float(ra) if ra.isdigit() else _seconds_until_httpdate(ra)
    remaining = int(resp.headers.get("X-RateLimit-Remaining", "1"))
    if remaining <= 1:
        reset = float(resp.headers.get("X-RateLimit-Reset", "0"))
        return max(0.0, reset - time.time())  # or reset directly if delta-seconds
    return 0.0

For sustained pulls, gate every request through a token-bucket sized to the documented budget (e.g. 60 req/min → refill 1 token/sec, capacity 60) so bursts smooth out instead of slamming the wall.

Pagination — loop to exhaustion

There is no single best strategy; the vendor chose one and you follow it. The universal rule: loop until the API says there is no next page — never a fixed page count. A hardcoded for page in range(10) silently drops everything after page 10.

(graphql.org/learn/pagination + pagination pattern guides, accessed 2026-06-02.)

Expose results as a generator / async-iterator so callers stream instead of buffering everything:

pythonCopy code
def iter_records(client):
    cursor = None
    while True:
        page = client.get("/items", params={"cursor": cursor, "limit": 100})
        body = page.json()
        yield from body["data"]
        cursor = body.get("next_cursor")
        if not cursor:           # exhaustion signal, not a counter
            return

Full code for every style — offset, keyset, Link-header parsing, GraphQL Relay connection walking, in Python and TS, plus dedup-on-overlap notes — is in references/pagination.md.

Putting it together

A minimal connector wires all four pillars plus env config and structured logging.

pythonCopy code
# connector.py — Python: httpx + tenacity
import os, logging, httpx
from tenacity import retry, stop_after_attempt, wait_exponential_jitter, retry_if_exception_type

log = logging.getLogger("connector")
TOKEN = os.environ["VENDOR_API_TOKEN"]          # from env, never hardcoded

class Transient(Exception): ...

@retry(stop=stop_after_attempt(5),
       wait=wait_exponential_jitter(initial=0.5, max=20),
       retry=retry_if_exception_type(Transient))
def _request(client, method, path, **kw):
    r = client.request(method, path, timeout=10.0, **kw)   # per-request timeout
    log.info("req id=%s %s status=%s", r.headers.get("x-request-id"), path, r.status_code)
    if r.status_code in (429, 502, 503, 504):
        raise Transient(r.status_code)
    r.raise_for_status()
    return r

def client():
    return httpx.Client(base_url="https://api.vendor.com/v2",
                        headers={"Authorization": f"Bearer {TOKEN}"})  # header, not query

typescriptCopy code
// connector.ts — Node/TS: global fetch (undici) + bounded retry
const TOKEN = process.env.VENDOR_API_TOKEN!;          // from env, never hardcoded
const RETRYABLE = new Set([429, 502, 503, 504]);

export async function request(path: string, init: RequestInit = {}, attempt = 0): Promise<Response> {
  const res = await fetch(`https://api.vendor.com/v2${path}`, {
    ...init,
    headers: { Authorization: `Bearer ${TOKEN}`, ...init.headers },
    signal: AbortSignal.timeout(10_000),              // per-request timeout
  });
  console.info(JSON.stringify({ id: res.headers.get("x-request-id"), path, status: res.status, attempt }));
  if (RETRYABLE.has(res.status) && attempt < 4) {
    const ra = Number(res.headers.get("retry-after"));
    const wait = Number.isFinite(ra) && ra > 0 ? ra * 1000 : Math.min(20_000, 500 * 2 ** attempt) + Math.random() * 500;
    await new Promise((r) => setTimeout(r, wait));
    return request(path, init, attempt + 1);
  }
  return res;                                          // caller checks res.ok / paginates
}

Anti-patterns

Verify

Run scripts/verify.sh over the connector you write: it greps for hardcoded secrets, asserts a retry mechanism and a pagination loop and a request timeout exist, and flags localStorage token storage or plaintext token logging. It is a structure linter (read-only), not a behavior test — exit 0 on a clean target.