Download the PHP package songhieu/monitor-track-laravel without Composer

monitor-track for Laravel

songhieu/monitor-track-laravel sends structured events to monitor-track App Logs. Plain log lines can't carry this data: exceptions with real stack frames, exact queue job runs, scheduled task runs that register themselves, queue worker heartbeats, and slow or N+1 database queries with the line of your code that ran them. The wire format is described in docs/sdk-spec.md.

Requires PHP 8.0+ and Laravel 9, 10, 11, 12 or 13 (Monolog 2 or 3). Every combination is tested in CI on PHP 8.0–8.5, including the lowest supported versions. Laravel 9, 10 and 11 no longer receive security fixes; Composer 2.10+ refuses to install them while they have open advisories — upgrading is the real fix.

Install

On Kubernetes, run the test inside a pod with --pod-log:

kubectl exec gives the command its own stderr, which is not the container log the collector reads; --pod-log writes to the container's stderr (/proc/1/fd/2) instead. It must run as the same user as the container's main process, which is what kubectl exec does by default.

Package discovery registers the service provider and the MonitorTrack facade. The default transport writes one JSON line per event to stderr. The monitor-track collector already reads pod logs, so on Kubernetes you usually don't need to configure anything else — no key, and no network call from the app.

Outside Kubernetes (a VM, shared hosting), create an HTTP push log source in monitor-track to get a key (mt_src_…, shown once) and send the events straight to the ingest service:

To publish the config file (optional):

Configuration

These MT_* variables are the same in the Go and Python SDKs.

With MT_TRANSPORT=http and no MT_ENDPOINT, the SDK falls back to stream. config/monitor-track.php also has capture.exceptions|queue|schedule|queries switches and the http timeouts.

What is captured automatically

The task name is the task's description if it has one. Otherwise it is the artisan command without the PHP binary (for example invoices:send-reminders). Closures with no ->name() are named closure:<file>:<line>.

Slow queries and N+1

The SDK listens to Laravel's QueryExecuted event and reports two problems as query events:

• N+1: the same statement runs MT_N_PLUS_ONE (10) times or more in one request, queued job, artisan command or scheduled task. Usually this is a relation lazy-loaded in a loop. The event has the number of runs, their total and slowest time, and the stack of your code that ran the statement: frames[0] is your line, such as the controller loop or the @foreach in a compiled Blade view.
• Slow query: a query that takes MT_SLOW_QUERY_MS (500 ms) or longer. Slow runs of the same statement from the same line are added up per request or job.

Statements are compared by shape: in (?, ?, ?) lists become in (?), and string and number literals become ?. The SDK never reads the bindings, so no values leave the app, and the SQL sent is at most 2000 bytes. A statement repeated only inside framework or vendor code, with none of your code on the stack (the migrator, the queue and cache drivers), is not reported as N+1.

Findings are sent when the unit of work ends: after the response for a request, and when the job, command or scheduled task finishes. A queue worker (queue:work, Horizon) and the other long-running commands (schedule:work, octane:*, reverb:start, pulse:*) are not a unit themselves. Each job they run is one, and the worker's own polling queries are ignored. Add your own daemon commands to that list with MT_LONG_RUNNING_COMMANDS=distribution:*,reports:daemon (names or * patterns); otherwise a command that loops for hours is one unit, and a statement it repeats over its lifetime is reported as N+1 when it exits. Counters start empty for every job and request, so nothing builds up in a long-running worker or an Octane process.

Events are samples. The SDK sends at most one event per statement and call site per MT_QUERY_THROTTLE_SECONDS (60) per server. An N+1 endpoint serving 100 requests a second sends one event a minute, not 100 a second. The count in an event is for that one request or job. The window is kept in marker files in the system temp dir (mt-q-*), so it works across PHP-FPM requests and between FPM and the queue workers. If the temp dir is not writable, 1% of the events are sent instead.

To tune it or turn it off:

With both set to 0, or 'capture' => ['queries' => false] in config/monitor-track.php, the SDK registers no query listener at all.

Overhead. Each query costs about 0.5 µs in the SDK: one normalization, remembered for repeated SQL, and one counter update. Laravel's own event dispatch adds about 1 µs. The SDK only takes a backtrace (under 0.1 ms) when a statement reaches the N+1 threshold, once per request or job, and for each slow query. Memory per request or job is bounded to 1000 distinct statements and 100 findings of each kind. Events are built and throttled when the request or job ends, which is after the response under FPM.

Logging channel

The provider adds a monitor-track channel. Its records become log events, or exception events when the context has an exception:

Pick the setup that matches your transport:

• stream transport (default): replace stderr with monitor-track, for example LOG_STACK=monitor-track (Laravel 11+) or 'channels' => ['monitor-track'] in the stack channel of config/logging.php (Laravel 9 and 10, which have no LOG_STACK). Both channels write to stderr, so LOG_STACK=stderr,monitor-track would print every line twice.
• file / http transport: LOG_STACK=stderr,monitor-track (Laravel 11+) or ['single', 'monitor-track'] (Laravel 9/10) is fine. The first channel stays human-readable, and the structured copy goes to the file or the ingest API. If the backend also reads this pod's stdout, you will see both copies.

Each exception is sent once, even when both the exception handler and a log call see it.

You can also attach the handler yourself: new MonitorTrack\Monolog\Handler(app(MonitorTrack\Client::class), 'warning').

Manual API

trackJob and trackCron record the exception and rethrow it unchanged.

In your application's tests, MonitorTrack::fake() routes events to memory:

Transports

• stream (default): one fwrite of one line to php://stderr or php://stdout. There is no network and no buffering. PHP-FPM: FPM only forwards worker stderr when it is configured to. In the pool config set:

  and in the global php-fpm.conf: log_limit = 32768 (the default of 1024 splits long lines).

• file: appends to MT_FILE with an exclusive flock, one write per line. The file is opened for each write, so logrotate works. Ship the file with Vector or Fluent Bit.
• http: events are buffered in memory (at most MT_QUEUE_SIZE, the newest are dropped when full). They are sent as gzip NDJSON to {MT_ENDPOINT}/api/v1/logs/ingest from app()->terminating(), which runs after the response is sent under FPM and when an artisan command ends. A console process also sends when due (2 s or 500 events) as it records events, so a daemon command delivers while it runs. Queue workers instead send every 2 s or 500 events from the Looping event, between jobs, and at WorkerStopping; Horizon's master and supervisor processes send from their once-a-second loop. Octane workers send when due, see Laravel Octane. A shutdown function is the last resort. Each request has a 0.5 s connect timeout and a 1 s total timeout. A connection error gets one retry after 200 ms. Any non-2xx response drops the batch. When the ingest can't be reached or answers 429 / 502–504, a long-running process stops trying for 5 s, doubling up to 60 s: events keep being buffered (up to MT_QUEUE_SIZE) and go out when it is back.

Laravel Octane

The SDK works under Octane (Swoole, RoadRunner) without configuration:

• Per-request context. Octane serves every request from a fresh clone of the booted application. The SDK reads the route, user and trace id from the current container at the time of each event, so every event carries its own request's route, user_id and trace_id, and exceptions are reported from every request, whichever copy of the exception handler it resolves.
• Queries. N+1 and slow-query counters start empty for every request (and every Octane task and tick), so nothing adds up across requests.
• http transport. An Octane worker serves many requests, so it does not POST after each one. After a request, the buffer is sent when it is due: 2 s since the last send, or 500 events. On Swoole, a one-shot timer sends what is left about 2 s after the last request; it fires between requests, never during one. Task workers send after their tasks and ticks, and every worker sends everything at WorkerStopping (max requests reached, octane:reload). A worker is detected by the LARAVEL_OCTANE variable Octane sets for its server.

Octane (1.x and 2.x) stops Swoole with SIGKILL (octane:stop, or SIGTERM to octane:start, which is what supervisord and Kubernetes send), so no PHP code runs at that point: up to the last ~2 s of events of each worker can be lost when the server is stopped. The same goes for a worker killed for exceeding Octane's max_execution_time. With RoadRunner there is no timer: what a request leaves in the buffer goes with a later request or when the worker stops.

With the stream transport, worker output passes through octane:start, which re-encodes JSON lines before printing them. Prefer the http or file transport under Octane.

Why this never slows your app

• No network on the request or job path. The default transport is a single local fwrite. The http transport only buffers during a request and sends after the response. Workers send between jobs.
• Fail-open. Every entry point, listener, transport and the service provider itself catches \Throwable and never rethrows. Failures are counted (MonitorTrack::stats()['errors']) and never raised. A broken config disables the SDK. It never breaks boot.
• Bounded memory. The http buffer holds at most MT_QUEUE_SIZE lines (1000), and the newest are dropped when it is full. Context is walked at most 5 levels deep and 100 items per level.
• Bounded time. Each http request has a 0.5 s connect and 1 s total timeout, with one quick retry and no loops. flush() stops at its budget. After a failed send, long-running processes (queue workers, Octane) back off for 5–60 s instead of waiting for the timeout at every flush.
• Bounded size. Lines stay under 16 KiB where possible and never exceed 32 KiB. The message, output and context are truncated first, and the event is dropped only as a last resort. Output is always a single line.
• No side effects. The SDK reads the user id only if the guard already loaded the user. It never resolves services the app hasn't resolved. It never logs through your logger.
• Sampling. MT_SAMPLE_RATE thins log events. Exceptions, jobs, cron runs and heartbeats are always kept. query events are throttled to one per statement and call site per minute per server.
• Cheap query hooks. About 0.5 µs per query in the SDK. Stacks are only captured when a threshold is crossed, and findings are sent after the response.

Known limits

• The PHP worker is single-threaded, so no heartbeat is sent while a job runs. A busy heartbeat is sent at job start when one is due, and heartbeats resume between jobs. Jobs that run longer than 3 × MT_HEARTBEAT_SECONDS go quiet until they finish.
• runInBackground() tasks report success or failure from the schedule:finish process. That process doesn't know the start time, so these events carry no duration_ms. Laravel sends the output of that process to /dev/null (and the task's own output to /dev/null or its appendOutputTo() file), so with the stream transport these events are lost: use the http or file transport for a scheduler with background tasks.
• A daemon command that records no events for a while holds what it buffered until its next event or its exit: the check runs when an event is recorded. Call MonitorTrack::flushIfDue() in its loop if that matters.
• N+1 counts runs of the same statement per request or job, wherever they come from. The call site reported is the one of the run that reached the threshold.
• An N+1 in a Blade template points at the compiled view (storage/framework/views/….php), not at the .blade.php file.