Troubleshooting Image API Connection Drops

{
  "status_code": 500,
  "error": {
    "message": "write tcp 10.0.0.1:443->203.0.113.5:52310: write: connection reset by peer",
    "type": "shell_api_error",
    "code": "write_response_body_failed"
  }
}

python3 -c "import ssl; print(ssl.OPENSSL_VERSION)"

NotOpenSSLWarning: urllib3 v2 only supports OpenSSL 1.1.1+,
currently the 'ssl' module is compiled with 'LibreSSL 2.8.3'

# macOS: build a virtualenv on Homebrew's Python
brew install [email protected]
python3.13 -m venv venv
venv/bin/pip install requests pillow
venv/bin/python -c "import ssl; print(ssl.OPENSSL_VERSION)"   # should print OpenSSL 3.x

# Inspect current values
sysctl net.ipv4.tcp_keepalive_time net.ipv4.tcp_keepalive_intvl net.ipv4.tcp_keepalive_probes

# Adjust temporarily (inside containers this needs --sysctl or privileges;
# in production prefer sysctl.d or setting SO_KEEPALIVE in the app)
sudo sysctl -w net.ipv4.tcp_keepalive_time=60
sudo sysctl -w net.ipv4.tcp_keepalive_intvl=15

ip link show            # inspect the container interface MTU
# Probe the real usable MTU with do-not-fragment packets
ping -M do -s 1400 api.apiyi.com

# Was the container OOM-killed?
dmesg -T | grep -i -E "oom|killed process"
kubectl describe pod <pod> | grep -A3 "Last State"   # look for OOMKilled

env | grep -i -E "proxy|no_proxy"

echo "--- proxy vars ---";  env | grep -i proxy || echo "none"
echo "--- keepalive ---";   sysctl net.ipv4.tcp_keepalive_time
echo "--- MTU ---";         ip link show | grep mtu
echo "--- fd limit ---";    ulimit -n
echo "--- DNS ---";         getent hosts api.apiyi.com
echo "--- reachability and timing ---"
curl -sS -o /dev/null -w 'connect=%{time_connect} tls=%{time_appconnect} ttfb=%{time_starttransfer} ip=%{remote_ip}\n' \
  https://api.apiyi.com/v1/models -H "Authorization: Bearer $KEY"

import { Agent, setGlobalDispatcher } from "undici";
import OpenAI from "openai";

// Image endpoints mean long silences plus MB-scale bodies — widen all three
setGlobalDispatcher(new Agent({
  connect: { timeout: 30_000 },   // connect in 30s; the default is only 10s
  headersTimeout: 300_000,        // first byte within 300s
  bodyTimeout: 300_000,           // inter-chunk gap up to 300s
}));

const client = new OpenAI({
  apiKey: process.env.APIYI_API_KEY,
  baseURL: "https://api.apiyi.com/v1",
  timeout: 300_000,   // total timeout — a different layer; set both
  maxRetries: 0,      // critical, see below
});

const agent = new Agent({ pipelining: 0, keepAliveTimeout: 1_000 });
// errors disappear ⇒ it was dead-connection reuse

curl -sS -o /tmp/out.json --trace-time \
  -w '\nconnect=%{time_connect} tls=%{time_appconnect} ttfb=%{time_starttransfer} total=%{time_total} bytes=%{size_download} code=%{http_code} ip=%{remote_ip}\n' \
  -H "Authorization: Bearer $KEY" -H 'Content-Type: application/json' \
  -d '{"model":"gpt-image-2-all","prompt":"a red cube on a white table"}' \
  https://api.apiyi.com/v1/images/generations

import time
import requests

TRANSPORT_ERRORS = (
    requests.exceptions.SSLError,
    requests.exceptions.ConnectionError,
    requests.exceptions.ChunkedEncodingError,
    requests.exceptions.ReadTimeout,
)

def call_image_api(url, headers, body, timeout=360, retries=2):
    """Retry transport-level failures up to `retries` times; never retry HTTP
    4xx/5xx — hand those straight back to the caller.

    Note: each retry may be a newly billed request, so keep `retries` small.
    """
    attempts = []
    for i in range(retries + 1):
        try:
            resp = requests.post(url, headers=headers, json=body,
                                 stream=True, timeout=(10, timeout))
            raw = b"".join(resp.iter_content(chunk_size=8192))
            attempts.append({"attempt": i + 1, "status": resp.status_code})
            return resp.status_code, raw, attempts      # 4xx/5xx included
        except TRANSPORT_ERRORS as e:
            attempts.append({"attempt": i + 1, "error": repr(e)})
            if i == retries:
                raise
            time.sleep(2 + 3 * i)                       # back off 2s, then 5s

resp = requests.post(url, headers=headers, json=body, stream=True, timeout=(10, 360))
chunks, total = [], 0
for chunk in resp.iter_content(chunk_size=8192):
    total += len(chunk)
    chunks.append(chunk)
raw = b"".join(chunks)
# total far below Content-Length  => the transfer broke midway
# total complete but connection stays open => upstream omitted the chunked
#   terminator, which is a channel-side problem