Five PHP-FPM workers capped an entire ride-hailing API at roughly 80 requests a second

Ninety-seven percent of a 62ms ping is framework overhead by the author's own loopback measurement, which leaves roughly 1.9ms for the work the endpoint exists to do ^[19][20][25]. That ratio is the case for a Go presence service, and the design was drawn: a WebSocket RPC writing straight to Redis, keeping GPS out of the PHP HTTP kernel ^[18]. The driver app posts a position every few seconds, and the post calls that endpoint the highest-volume thing in the system by an order of magnitude ^[17].

Then the config file. `pm.max_children` was 5, the stock PHP-FPM default, never tuned, on an eight-core box, which is five concurrent requests for the whole API ^[21]. At a 62ms median that is roughly 80 req/s ^[22]. The ceiling matters less than the queueing behind it: five slow I/O-bound requests, a routing call and a report query among them, stall every other endpoint, and a rider's booking queues behind GPS pings ^[23]. The value went to 24 ^[24]. Same 62ms median, 24 workers, and the ceiling is about 387 req/s, roughly 4.8 times what five allowed ^[26]. The Go sidecar was built and ships off by default ^[8]. The post describes itself as "a set of architecture decisions I can point at a number for, and a few that a number reversed" ^[28].

The separation runs along change rate ^[3]. Pricing is scoped: the pickup resolves to a service area, that area is recorded on the trip, and from then on it decides the fare table, the surge, the match radius and how long a driver has to answer an offer ^[10]. That is what makes one Laravel deployment per country workable, each with its own database and its own worker containers against a shared Postgres instance ^[5][4].

Matching runs a Redis `GEORADIUS ... ASC` around the pickup and walks the result first-eligible-wins ^[11]. Drivers set their own willingness to travel, so the scan goes to the platform ceiling and each candidate is then filtered against the radius that driver chose ^[12]. A wider scan surfaces work a driver could not previously see; it never lets them jump ahead of someone closer ^[12]. Accepting an offer locks only the target trip row, which the post says is not enough, so Postgres carries a partial unique index, `one_active_trip_per_driver` on `trips (captain_id)`, restricted to the four in-flight statuses ^[14].

Presence can leave PHP. Pricing has to stay. Every GPS fix writes a route point, and completion reads that trail to finalize the fare ^[15]. "The trail is the fare," the author wrote, giving that as the reason it stays in PHP ^[16].

The map stack is self-hosted: OSRM for routing, Photon for geocoding, TileServer GL for imagery ^[9]. The per-geocode price of the hosted services those replace is not in the published text ^[27]. For the 80 req/s number to mean anything on another deployment, two things have to hold: an untuned `pm.max_children = 5`, and a median per request near 62ms. The measurement was taken over loopback on a system in UAT, and the post says nothing in it describes real scale ^[2][19].