Rollbar

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me an error tracker that replaces Rollbar.

This is a small service with one genuinely difficult part, and everything below is arranged around it.

STACK
- Node 20+ with Fastify
- SQLite through better-sqlite3, WAL mode, FTS5 over the messages
- Caddy in front

THE DATA MODEL
- groups: id, fingerprint, title, culprit, level, first_seen, last_seen, count, status, assignee, resolved_in_release
- occurrences: id, group_id, project_id, release, environment, message, stack_json, context_json, person_hash, at
- releases: id, project_id, version, deployed_at
- Occurrences pruned on a retention; groups kept forever. **The count and the first-seen date are the valuable part and they cost nothing to keep**

GROUPING, WHICH IS THE WHOLE PRODUCT
- Fingerprint from the exception type plus the top frames of the stack that belong to your own code — skip framework and node_modules frames, or every error in the framework becomes one group
- **Normalise the variable parts out**: numbers in paths, hex ids, quoted values, memory addresses. 'User 12345 not found' and 'User 67890 not found' are one bug
- Line numbers move when a file is edited. Fingerprint on function names and file paths, not on line numbers, or a release with a blank line at the top creates a hundred new groups
- Let a project override the fingerprint with a rule, because you will always find one case the generic algorithm splits

SOURCE MAPS, THE OTHER HARD PART
- Upload maps at deploy time keyed by release; resolve frames on ingest and store both the raw and the resolved
- If the map is missing say so on the frame rather than showing minified rubbish silently

INGEST THAT CANNOT TAKE THE SITE DOWN
- Accept, write, respond. Never process synchronously
- **A rate limit per group, not just per project.** An error in a hot loop generating a million events in a minute is the normal failure of these systems, and it should cost you a counter increment rather than a disk
- A payload size cap and a stack depth cap

BEING TOLD
- Alert on a new group, on a group returning after being resolved, and on a spike against the group's own baseline. Not on every occurrence
- **A daily digest, not a message per error.** An alerting system people mute is worse than none
- Resolve a group in a release; if it occurs in a later release, reopen it and say so

CONTEXT WORTH HAVING
- Release, environment, request path, a hashed person id, and a short breadcrumb trail
- **Scrub before storing**: passwords, tokens, card numbers, authorisation headers. Deny by default on anything that looks like a secret, and test it

WHAT MATTERS MOST
Fingerprinting and the per-group rate limit. Bad grouping makes it unreadable; no rate limit makes it a liability.