Grammarly Observability

Overview

Grammarly API integrations process user text through scoring, AI rewriting, and plagiarism endpoints where latency and accuracy directly affect user experience. Monitor text check response times, suggestion quality signals, API error rates, and token consumption to stay within rate limits. Catching degradation early prevents users from seeing stale suggestions or silent failures in real-time editing flows.

Key Metrics

Instrumentation

async function trackGrammarlyCall(api: 'score' | 'ai' | 'plagiarism', textLen: number, fn: () => Promise<any>) {
  const start = Date.now();
  try {
    const result = await fn();
    metrics.histogram('grammarly.api.latency', Date.now() - start, { api });
    metrics.increment('grammarly.api.calls', { api });
    metrics.gauge('grammarly.text.length', textLen, { api });
    return result;
  } catch (err) {
    metrics.increment('grammarly.api.errors', { api, status: err.status });
    throw err;
  }
}

Health Check Dashboard

async function grammarlyHealth(): Promise<Record<string, string>> {
  const latencyP95 = await metrics.query('grammarly.api.latency', 'p95', '5m');
  const errorRate = await metrics.query('grammarly.api.error_rate', 'avg', '5m');
  const quotaUsed = await grammarlyAdmin.getQuotaUsage();
  return {
    api_latency: latencyP95 < 300 ? 'healthy' : 'slow',
    error_rate: errorRate < 0.01 ? 'healthy' : 'degraded',
    quota: quotaUsed < 0.8 ? 'healthy' : 'at_risk',
  };
}

Alerting Rules

const alerts = [
  { metric: 'grammarly.api.latency_p95', condition: '> 800ms', window: '10m', severity: 'warning' },
  { metric: 'grammarly.api.error_rate', condition: '> 0.05', window: '5m', severity: 'critical' },
  { metric: 'grammarly.quota.daily_pct', condition: '> 0.90', window: '1h', severity: 'warning' },
  { metric: 'grammarly.ai.throughput', condition: 'drop > 30%', window: '15m', severity: 'critical' },
];

Structured Logging

function logGrammarlyEvent(api: string, data: Record<string, any>) {
  console.log(JSON.stringify({
    service: 'grammarly', api,
    duration_ms: data.latency, status: data.status,
    text_length: data.textLen, suggestion_count: data.suggestions,
    // Never log user text content — only metadata
    timestamp: new Date().toISOString(),
  }));
}

Error Handling

Prerequisites

• Data-minimization policy for logs/traces, metric owners, alert routes, retention limits, and an approved restricted evidence destination.
• Synthetic probes for availability, quota, and authorization/consent behavior plus correlation IDs that contain no text or identity.

Instructions

1. Instrument aggregate request, error, latency, quota, queue, and retention-policy metrics at the integration boundary.
2. Redact event fields before export, bound payload capture, and fail closed when redaction cannot be proven.
3. Alert on sustained failure, quota saturation, access/retention policy difference, and evidence-pipeline failure with a named runbook owner.
4. Test alerts with synthetic events, verify destination and retention, then remove temporary test artifacts.
5. Review dashboards after any credential, environment, or client change and retain only the minimum reconstruction evidence.

Output

Produce an observability receipt with dashboard/alert revisions, synthetic results, owner, destination, retention, correlation ID, and remediation/rollback status. Raw text and credentials are prohibited.

Examples

dashboard=writing-r11; alert=quota-lag; probe=pass; destination=restricted-ops; retention=14d; rollback=alert-r10 is a safe alert-test result.

Resources

• Grammarly Developer Portal

Next Steps

See grammarly-incident-runbook.