You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit introduces a new `Connection::graceful_shutdown_with_config`
method that gives users control over the HTTP/1 graceful process.
Before this commit, if a graceful shutdown was initiated on an inactive
connection hyper would immediately close it.
As of this commit the `GracefulShutdownConfig::first_byte_read_timeout`
method can be used to give inactive connections a grace period where,
should the server begin receiving bytes from them before the deadline,
the request will be processed.
## HTTP/2 Graceful Shutdowns
This commit does not modify hyper's HTTP/2 graceful shutdown process.
`hyper` already uses the HTTP/2 `GOAWAY` frame, meaning that `hyper`
already gives inactive connections a brief period during which they can
transmit their final requests.
Note that while this commit enables slightly more graceful shutdowns for
HTTP/1 servers, HTTP/2 graceful shutdowns are still superior.
HTTP/2's `GOAWAY` frame allows the server to finish processing a last
batch of multiple incoming requests from the client, whereas the new
graceful shutdown configuration in this commit only allows the server to
wait for one final incoming request to be received.
This limitations stems from a limitation in HTTP/1, where there is
nothing like the `GOAWAY` frame that can be used to coordinate the
graceful shutdown process with the client in the face of multiple
concurrent incoming requests.
Instead for HTTP/1 connections `hyper` gracefully shuts down by
disabling Keep-Alive, at which point the server will only receive at
most one new request, even if the client has multiple requests that are
moments from reaching the server.
## Motivating Use Case
I'm working on a server that is being designed to handle a large amount
of traffic from a large number of clients.
It is expected that many clients will open many TCP connections with the
server every second.
As a server receives more traffic it becomes increasingly likely that at
the point that it begins gracefully shutting down there are connections
that were just opened, but the client's bytes have not yet been seen by
the server.
Before this commit, calling `Connection::graceful_shutdown` on such a
freshly opened HTTP/1 connection will immediately close it. This means
that the client will get an error, despite the server having been
perfectly capable of handling a final request before closing the
connection.
This commit solves this problem for HTTP/1 clients that tend to send one
request at a time.
By setting a `GracefulShutdownConfig::first_byte_read_timeout` of, say,
1 second, the server will wait a moment to see if any of the client's
bytes have been received.
During this period, the client will have been informed that Keep-Alive
is now disabled, meaning that at most one more request will be
processed.
Clients that have multiple in-flight requests that have not yet reached
the server will have at most one of those requests handled, even if all
of them reach the server before the `first_byte_read_timeout`.
This is a limitation of HTTP/1.
## Work to do in other Crates
#### hyper-util
To expose this to users that use `hyper-util`, a method should be added
to `hyper-util`'s `Connection` type.
This new `hyper-util Connection::graceful_shutdown_with_config` method
would expose a `http1_first_byte_read_timeout` method that would lead
`hyper-util` to set `hyper
GracefulShutdownConfig::first_byte_read_timeout`.
---
Closeshyperium#3792
0 commit comments