DEV Community

Cover image for OkHttp connectTimeout(10, SECONDS) Was Client Folklore — We Changed Timeouts Live Mid-Outage (Java)
Moshe Avdiel
Moshe Avdiel

Posted on • Originally published at github.com

OkHttp connectTimeout(10, SECONDS) Was Client Folklore — We Changed Timeouts Live Mid-Outage (Java)

Thursday 03:19. Payments calls a card network over OkHttp. Their edge is slow — not dead. Your client was built once:

OkHttpClient client = new OkHttpClient.Builder()
    .connectTimeout(10, TimeUnit.SECONDS)
    .readTimeout(30, TimeUnit.SECONDS)
    .writeTimeout(30, TimeUnit.SECONDS)
    .build();
Enter fullscreen mode Exit fullscreen mode

P99 climbs past 30s. Threads pile on readTimeout. Someone wants 8s fail-fast to trip the circuit breaker earlier. That number is trapped in a final client from startup.

"OkHttp timeouts are client construction. New values need a release."

The Aha: connect/read/write timeouts and call timeouts are dependency health policy — operational, not folklore. Kiponos.io holds them live; you either rebuild a lightweight client on change or set per-call timeouts from local gets.

Belief vs production

Belief Reality
30s read is "safe" Safe for you, lethal for thread pools
One client forever Partner SLOs change by region and hour
Retry in code constants Retries without live budgets amplify outages

Kiponos shape 

http_ops/
  card_network/
    connect_timeout_ms: 3000
    read_timeout_ms: 8000
    write_timeout_ms: 8000
    call_timeout_ms: 10000
    max_retries: 1
    retry_backoff_ms: 200
Enter fullscreen mode Exit fullscreen mode

Profile: ['payments']['v3']['prod']['base']. Hot path: local getInt — zero network.

Architecture

Architecture diagram

Integration 

@Component
public class LiveCardNetworkHttp {

    private final Kiponos kiponos;
    private final AtomicReference<OkHttpClient> client = new AtomicReference<>();

    public LiveCardNetworkHttp(Kiponos kiponos) {
        this.kiponos = kiponos;
        client.set(build());
        kiponos.afterValueChanged(ch -> {
            if (ch.path().startsWith("http_ops/card_network")) {
                client.set(build());
            }
        });
    }

    private OkHttpClient build() {
        var h = kiponos.path("http_ops", "card_network");
        return new OkHttpClient.Builder()
            .connectTimeout(h.getLong("connect_timeout_ms", 3000), TimeUnit.MILLISECONDS)
            .readTimeout(h.getLong("read_timeout_ms", 8000), TimeUnit.MILLISECONDS)
            .writeTimeout(h.getLong("write_timeout_ms", 8000), TimeUnit.MILLISECONDS)
            .callTimeout(h.getLong("call_timeout_ms", 10000), TimeUnit.MILLISECONDS)
            .build();
    }

    public String post(String url, RequestBody body) throws IOException {
        var h = kiponos.path("http_ops", "card_network");
        int attempts = 1 + h.getInt("max_retries", 1);
        IOException last = null;
        for (int i = 0; i < attempts; i++) {
            try {
                Request req = new Request.Builder().url(url).post(body).build();
                try (Response resp = client.get().newCall(req).execute()) {
                    return resp.body() != null ? resp.body().string() : "";
                }
            } catch (IOException e) {
                last = e;
                sleep(h.getLong("retry_backoff_ms", 200));
            }
        }
        throw last;
    }
}
Enter fullscreen mode Exit fullscreen mode

Per-call override without full rebuild:

int readMs = kiponos.path("http_ops", "card_network").getInt("read_timeout_ms", 8000);
Request req = new Request.Builder().url(url).get().build();
try (Response resp = client.get().newBuilder()
        .readTimeout(readMs, TimeUnit.MILLISECONDS)
        .build()
        .newCall(req)
        .execute()) {
    // ...
}
Enter fullscreen mode Exit fullscreen mode

Scenarios

Moment Frozen reflex Live
Slow partner Hope / deploy Drop read_timeout_ms to 5s
Regional blip Global 60s Profile-specific timeouts
Load test Hard-coded Hub loadtest tree

Pair with Feign timeouts and WebClient.

Before / after

Approach Change timeout mid-outage Hot path
final client Redeploy Frozen
Config server poll per call Possible Network RTT
Kiponos Delta + rebuild or per-call Local get

When not

Case Prefer
mTLS cert rotation Secret manager + restart policy
Base URL environment Deploy-time wiring
Protocol change HTTP→gRPC Architecture

Performance notes

  • Rebuild OkHttpClient on delta is cheap compared to an outage; prefer that over polling a config server on every call.
  • getInt / getLong remain O(1) local memory — safe inside retry loops.
  • Do not put secrets, base URLs, or mTLS material in the live tree; only operational floats.
  • Dispatcher thread pool sizes are a separate dial — pair with server-side thread articles when both sides thrash.

Getting started

  1. Move four timeout ints + retry budget into hub
  2. AtomicReference client rebuild on afterValueChanged
  3. Game day: inject latency, tighten read timeout from dashboard
  4. Keep base URL and certs out of the live tree
  5. Document which profile owns card-network vs warehouse HTTP so on-call never edits the wrong tree

Resources: github.com/kiponos-io/kiponos-io · GETTING-STARTED

Kiponos.io — HTTP client timeouts are dependency health policy, not folklore.

Top comments (0)