# Where your messages are kept

Queues run on the platform, not on machines of yours; this page says where messages are stored, what happens in the background, and which numbers are approximate.

## What runs where

No machine of yours is involved. A queue and its messages are stored in the platform's own database
in the queue's region, shared with other organizations' queues; the limits in [Quotas](quotas.md)
keep one organization's traffic from slowing the others. The queue API and the SQS endpoint are
two doors to the same service, so limits, delivery and dead-letter rules are the same through both.

## What happens in the background

Once a minute, the platform:

- removes messages whose retention has run out;
- moves messages that have used up their receives to their dead-letter queue, for every queue that
  has one, so the dead-letter counts are right even when nobody receives (a receive also does this
  for its own queue first);
- records each queue's available, in-flight and delayed counts for the
  [Messages over time](monitoring.md) chart.

A redrive runs inside the request that starts it, and answers when it has finished.

## Long polling

A waiting receive is woken at once by a send through the same server; otherwise it looks for new
messages every second until its wait is up. It stops when the client disconnects.

## Approximate numbers

- The checks for messages per queue, in-flight messages per queue and stored bytes per
  organization use counts that are up to a few seconds old. A burst can go slightly over a limit
  before requests are refused.
- The request rate and the open long polls are counted by each of the platform's servers on its
  own, so your organization can get somewhat more than the stated rate.
- Request counts for billing are collected every 10 seconds; requests counted by a server that
  stops in between are not billed.

## Related

- [Message delivery](message-delivery.md)
- [Monitoring](monitoring.md)
