Skip to content

Service level

This page is the service level agreement (SLA) for Tarno's REST API (https://api.tarno.cl) and MCP server (https://mcp.tarno.cl): what we commit to, how each figure is measured and what is left out. The figures are deliberately conservative: each one sits below what the service has actually measured (shown alongside), so the commitment can be met in a bad month too.

The consequences of a breach (credits or otherwise) are set by your organisation's contract; this page sets the objectives and how they are measured.

Architecture: one machine, off-site copies and a rehearsed recovery

Tarno runs on a single machine. There is no hot standby and no automatic failover: a machine failure is covered by recovering the service from backup, not by moving traffic to another one. That is why the commitments on this page are about monthly availability and recovery time, not redundancy.

  • Encrypted copies off the machine, in independent external storage: the corpus and the watch data once a day; the clients' API keys and organisations every hour. They are kept for 14 days.
  • Archive of the Estado Diario PDFs: the bulletins as INAPI published them are copied every day to the same external storage, the integrity of the whole copy is checked (every file, by size and checksum) and they are kept with no expiry. They are not encrypted: they are public documents.
  • Rehearsed recovery: the full restore of the service on a new machine, from the real copies, is written down step by step and timed (last rehearsal: 2026-10-04).

Commitments

Commitment Committed value Measured (reference)
Monthly availability of the REST API and the MCP 99.5% 99.93% over 30 days (as of 2026-09-30)
Server error responses (5xx) < 0.5% of the month's requests 0.034% over 30 days
Maximum data loss (RPO) 24 h (keys and organisations: 1 h) daily copy of the corpus and watch data; hourly for keys and organisations
Recovery time (RTO) 4 h rehearsed full restore: 11 min of technical work until serving with the clients' own keys, ≈ 28 min to full service
Maintenance within the declared window, announced 48 h ahead —

Availability

Percentage of the minutes of the calendar month (UTC) in which the REST API and the MCP answer requests. A minute counts as unavailable when the service does not answer or answers 503 to every request. Minutes covered by the exclusions, announced maintenance included, do not count as unavailable. 99.5% is at most about 3 h 40 min of unavailability in a 31-day month.

Server errors

Share of the month's requests that end in a 5xx code. 4xx responses do not count (invalid request, key without permission, 429 for exceeding the quota or the burst limit): they are correct answers to a request that cannot be served.

Maximum data loss (RPO)

If the service had to be recovered from backup, the most that can be lost is what was written since the last copy:

data RPO
Corpus (trademarks, annotations, parties, catalogs) 24 h; the next sync with INAPI also brings back what was published
Similarity watch (lists, items and detected hits) 24 h
Clients' API keys and organisations 1 h

After a recovery, GET /system/freshness says how recent the served corpus is.

Recovery time (RTO)

Time from the confirmation of a machine failure until the REST API and the MCP answer again with the clients' keys and full search (similarity search included). The clock runs within support hours; a failure outside them starts counting at the beginning of the next period. The technical part of the restore is measured (about 28 min to full service); the rest of the window is margin to prepare the new machine, its names and certificates.

Maintenance

  • Declared window: Sundays 02:00–06:00 (America/Santiago), at most 4 h per month.
  • Advance notice: at least 48 h ahead, to your organisation's technical contact.
  • Maintenance outside the window or without notice counts as unavailability.
  • The daily sync is not maintenance: INAPI data is loaded with the service running, without an outage. During the daily chain the API and the MCP keep answering.

Exclusions

The following do not count against the commitments:

  • announced maintenance within the declared window;
  • force majeure and outages of network or name providers outside Tarno;
  • INAPI unavailability or that of its sources: it affects the freshness of the corpus, not the availability of the API, which keeps serving the data already loaded;
  • use beyond quota or the burst limit (429), and invalid or unauthorised requests (4xx);
  • attacks or abuse that force access to be limited to protect the service.

Performance objectives

Latency objectives (p95, measured over the month) we run the service under. Unlike the commitments above, they are objectives: if one is exceeded, it is investigated and fixed.

Operation p95 objective Measured (load test, 6 concurrent clients)
Trademark record, annotations and catalogs < 1.5 s record 1.2 s · annotations 0.4 s
Search < 3 s 1.1 s
Similarity search (similarTo) < 5 s 1.4 s warm; 4.5–4.7 s cold for short queries
Reports < 6 s clearance report 3.2 s

Freshness cadence

Tarno keeps its neutral corpus current by running its INAPI ingestion pipelines on a managed cron. The cadence follows how INAPI itself publishes:

Source INAPI publication cadence Tarno sync
INAPI open data Daily Managed cron, Monday to Saturday (no sync on Sundays)
Estado-Diario Monday–Friday (Santiago) Managed cron, on publication days
INAPI record pages Continuous Managed cron, once a day within the daily chain (12:00 Chile)

There is no synchronization on Sundays (decision of 2026-10-10): the chain runs Monday to Saturday at 12:00 Chile, so on Sunday dataAsOf stays on Saturday. stale in GET /system/freshness does not count Sundays (with the default threshold; with an explicit thresholdHours, exact). On Monday the chain picks up everything accumulated. INAPI does not publish the Estado Diario on Sundays.

Freshness objective: what INAPI publishes on a business day, available before 15:00 (America/Santiago) that day, unless INAPI did not publish it in time. All cadences are stated in America/Santiago time. The per-source freshness at any moment is reported by GET /system/freshness (scope insights:read) — see the status page. Treat GET /system/freshness, not this document, as the live truth for how fresh the corpus is right now.

Support

Support is handled by the operator during business hours (Monday to Friday, America/Santiago, except Chilean public holidays):

  • There is no 24/7 support. A failure outside business hours is handled at the start of the next business period, and the RTO counts from there.
  • For issues, contact your operator with the requestId from the error envelope (it correlates your report with the server log).
  • Breaking-change and deprecation notices are published in the changelog with the migration window stated in the versioning policy.

Summary

Aspect Commitment
Availability 99.5% monthly; 5xx errors < 0.5% of requests.
Data RPO 24 h (keys and organisations 1 h); encrypted off-machine copies, 14 days.
Recovery RTO 4 h within support hours; rehearsed restore.
Maintenance Sundays 02:00–06:00 (Santiago), max 4 h/month, 48 h notice. The daily sync does not interrupt the service.
Architecture One machine, no hot standby; recovered from backup.
Freshness Open data Monday to Saturday (no sync on Sundays); Estado-Diario Mon–Fri (Santiago); live in GET /system/freshness.
Support Operator, Chilean business hours. No 24/7 support.