Module limits 

The session’s request limits.

RequestLimits is shared by every worker that asks devices anything: a global cap on how many requests a second leave across every device, and one gate per device, so a device never has more than one of our confirmed requests outstanding, whether the browser, the watch list, a write or a managed COV renewal sent it (#1791). Small devices often handle one transaction at a time.

The MCP server adds two layers (#2066): a rate per network, since a slow MS/TP trunk behind a router takes far fewer requests than the IP side (RequestLimits::with_network_rates), and a per-caller Account with its own TokenBucket (charged).

Structs

Account

One caller’s tally: the requests it sent and how long they waited for the limits, and the caller’s token bucket, if it has one.

RequestLimits

Spaces requests evenly: at most per_second a second across every job, the browser’s and the watch list’s alike. Each caller takes the next free slot and sleeps until it; the lock is held only to take the slot, never across the sleep. It also remembers when recent requests went, for the request-rate meter.

TokenBucket

A token bucket: at most burst requests at once, refilled at per_second. A caller that finds it empty waits for its token, which is booked as it starts waiting, so callers are served in turn and a dropped wait still spends its token.

Turn

A request’s turn at its device: the device’s gate, held until this is dropped, on every path out of the request (an answer, an error, or the request’s future dropped).

Constants

METER_WINDOW_SECS

How long, in seconds, the request-rate meter (RequestLimits::recent) remembers a request.

Functions

charged

Run future with every request it sends charged to account.

refund

Take back the last request this task’s call was charged for, when its send failed before anything left (the transport refused it).