-
Notifications
You must be signed in to change notification settings - Fork 273
http: Lazily build http/tcp stacks #681
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Conversation
Currently, the `DetectHttp` layer always builds both HTTP and TCP stacks, even though only one is needed. Furthermore, we lose the information about which protocol was detected--it's never passed into the http stack (and instead must be inferred from individual requests). This change modifies the `DetectHttp` layer to only build inner services as needed, after the protocol has been detected. Furthermore, outbound caching has been moved up from the tcp balancer to wrap the HTTP detection layer. This was necessary so that the tcp balancer could be built instantaneously as a NewService while still being cloneable. This means that, for each outbound target ip:port, we will reuse HTTP and TCP stacks. However, HTTP routing remains dependent on per-request metadata. This is expected to change in a followup.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
this change looks great to me --- i commented on a few minor nits.
|
||
#[derive(Copy, Clone, Debug)] | ||
pub struct FromMetadata; | ||
|
||
#[derive(Clone, Debug, Eq, PartialEq, Hash)] |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
nit/TIOLI: i imagine these could also be Copy
?
@@ -53,6 +59,11 @@ pub struct HttpEndpoint { | |||
pub concrete: HttpConcrete, | |||
} | |||
|
|||
#[derive(Clone, Debug, Eq, PartialEq, Hash)] |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
this can also be Copy
, if you wanted it to.
linkerd/app/outbound/src/lib.rs
Outdated
.push_on_response( | ||
svc::layers() | ||
.push(tcp::balance::layer(EWMA_DEFAULT_RTT, EWMA_DECAY)) | ||
.push(svc::layer::mk(tcp::Forward::new)) | ||
) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
nit: some kinda weird formatting going on here?
http1: Option<H::Service>, | ||
h2: Option<H::Service>, |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
nit, doesn't actually matter: naming these http1
and h2
is kinda bugging me.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It corresponds to the Version
variant
let http1 = if let Some(svc) = self.http1.clone() { | ||
svc | ||
} else { | ||
let svc = self | ||
.new_http | ||
.new_service((HttpVersion::Http1, self.target.clone())); | ||
self.http1 = Some(svc.clone()); | ||
svc | ||
}; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
nit, take it or leave it: get_or_insert_with
might be more concise here
let http1 = if let Some(svc) = self.http1.clone() { | |
svc | |
} else { | |
let svc = self | |
.new_http | |
.new_service((HttpVersion::Http1, self.target.clone())); | |
self.http1 = Some(svc.clone()); | |
svc | |
}; | |
let http1 = self | |
.http1 | |
.get_or_insert_with(|| { | |
self | |
.new_http | |
.new_service((HttpVersion::Http1, self.target.clone())) | |
}) | |
.clone(); |
(i tried to format the suggested code the way i imagine rustfmt would want it)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
That doesn't work due to borrows on self.
let h2 = if let Some(svc) = self.h2.clone() { | ||
svc | ||
} else { | ||
let svc = self | ||
.new_http | ||
.new_service((HttpVersion::H2, self.target.clone())); | ||
self.h2 = Some(svc.clone()); | ||
svc | ||
}; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
could use get_or_insert_with
here as well.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
That doesn't work due to borrows on self.
let tcp = if let Some(svc) = self.tcp.clone() { | ||
svc | ||
} else { | ||
let svc = self.new_tcp.new_service(self.target.clone()); | ||
self.tcp = Some(svc.clone()); | ||
svc | ||
}; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
...and here.
) | ||
.into_make_service() | ||
.spawn_buffer(buffer_capacity) | ||
.instrument(|_: &_| info_span!("tcp")) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
nit: do we still want the instrument layer here?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
the tcp scope is now set in detect-http. it's probably better to stop doing that and keep them in here
This release includes changes to TCP metrics to ensure that peer identities are encoded via the `client_id` and `server_id` labels. --- * outbound: Explicitly ignore the source address for tap (linkerd/linkerd2-proxy#680) * Update proxy-api and tonic (linkerd/linkerd2-proxy#682) * http: Lazily build http/tcp stacks (linkerd/linkerd2-proxy#681) * outbound: Remove required identity from HttpLogical (linkerd/linkerd2-proxy#683) * profiles: Expose the fully_qualified_name (linkerd/linkerd2-proxy#684) * request-filter: Support altering the request type (linkerd/linkerd2-proxy#685) * tracing: Set contexts in new_service/make_service (linkerd/linkerd2-proxy#686) * discover: Allow resolution streams to terminate (linkerd/linkerd2-proxy#689) * metrics: add peer identities to all TLS metric labels (linkerd/linkerd2-proxy#687) * outbound: Return a default endpoint on reject (linkerd/linkerd2-proxy#690) * Skip endpoint resolution when profile lookup is rejected (linkerd/linkerd2-proxy#691)
This release includes changes to TCP metrics to ensure that peer identities are encoded via the `client_id` and `server_id` labels. --- * outbound: Explicitly ignore the source address for tap (linkerd/linkerd2-proxy#680) * Update proxy-api and tonic (linkerd/linkerd2-proxy#682) * http: Lazily build http/tcp stacks (linkerd/linkerd2-proxy#681) * outbound: Remove required identity from HttpLogical (linkerd/linkerd2-proxy#683) * profiles: Expose the fully_qualified_name (linkerd/linkerd2-proxy#684) * request-filter: Support altering the request type (linkerd/linkerd2-proxy#685) * tracing: Set contexts in new_service/make_service (linkerd/linkerd2-proxy#686) * discover: Allow resolution streams to terminate (linkerd/linkerd2-proxy#689) * metrics: add peer identities to all TLS metric labels (linkerd/linkerd2-proxy#687) * outbound: Return a default endpoint on reject (linkerd/linkerd2-proxy#690) * Skip endpoint resolution when profile lookup is rejected (linkerd/linkerd2-proxy#691)
this commit updates our tower dependency from 0.4 to 0.5. note that this commit does not affect the `tower-service` and `tower-layer` crates, reëxported by `tower` itself. the `Service<T>` trait and the closely related `Layer<S>` trait have not been changed. the `tower` crate's utilities have changed in various ways, some of particular note for the linkerd2 proxy. see these items, excerpted from the tower changelog: - **retry**: **Breaking Change** `retry::Policy::retry` now accepts `&mut Req` and `&mut Res` instead of the previous mutable versions. This increases the flexibility of the retry policy. To update, update your method signature to include `mut` for both parameters. ([#584]) - **retry**: **Breaking Change** Change Policy to accept &mut self ([#681]) - **retry**: Add generic backoff utilities ([#685]) - **retry**: **Breaking Change** `Budget` is now a trait. This allows end-users to implement their own budget and bucket implementations. ([#703]) - **util**: **Breaking Change** `Either::A` and `Either::B` have been renamed `Either::Left` and `Either::Right`, respectively. ([#637]) - **util**: **Breaking Change** `Either` now requires its two services to have the same error type. ([#637]) - **util**: **Breaking Change** `Either` no longer implemenmts `Future`. ([#637]) - **buffer**: **Breaking Change** `Buffer<S, Request>` is now generic over `Buffer<Request, S::Future>.` ([#654]) the `Either` trait bounds are particularly impactful for us. * <tower-rs/tower@v0.4.x...master> * <https://github.com/tower-rs/tower/blob/master/tower/CHANGELOG.md> * <tower-rs/tower#815> * <tower-rs/tower#817> * <tower-rs/tower#818> this work is based upon #3504. for more information, see: * linkerd/linkerd2#8733 * #3504 Signed-off-by: katelyn martin <kate@buoyant.io>
this commit updates our tower dependency from 0.4 to 0.5. note that this commit does not affect the `tower-service` and `tower-layer` crates, reëxported by `tower` itself. the `Service<T>` trait and the closely related `Layer<S>` trait have not been changed. the `tower` crate's utilities have changed in various ways, some of particular note for the linkerd2 proxy. see these items, excerpted from the tower changelog: - **retry**: **Breaking Change** `retry::Policy::retry` now accepts `&mut Req` and `&mut Res` instead of the previous mutable versions. This increases the flexibility of the retry policy. To update, update your method signature to include `mut` for both parameters. ([#584]) - **retry**: **Breaking Change** Change Policy to accept &mut self ([#681]) - **retry**: **Breaking Change** `Budget` is now a trait. This allows end-users to implement their own budget and bucket implementations. ([#703]) - **util**: **Breaking Change** `Either::A` and `Either::B` have been renamed `Either::Left` and `Either::Right`, respectively. ([#637]) - **util**: **Breaking Change** `Either` now requires its two services to have the same error type. ([#637]) - **util**: **Breaking Change** `Either` no longer implemenmts `Future`. ([#637]) - **buffer**: **Breaking Change** `Buffer<S, Request>` is now generic over `Buffer<Request, S::Future>.` ([#654]) see: * <tower-rs/tower#584> * <tower-rs/tower#681> * <tower-rs/tower#703> * <tower-rs/tower#637> * <tower-rs/tower#654> the `Either` trait bounds are particularly impactful for us. because this runs counter to how we treat errors (skewing towards boxed errors, in general), we temporarily vendor a version of `Either` from the 0.4 release, whose variants have been renamed to match the 0.5 interface. updating to box the inner `A` and `B` services' errors, so we satiate the new `A::Error = B::Error` bounds, can be addressed as a follow-on. that's intentionally left as a separate change, due to the net size of our patchset between this branch and #3504. * <tower-rs/tower@v0.4.x...master> * <https://github.com/tower-rs/tower/blob/master/tower/CHANGELOG.md> this work is based upon #3504. for more information, see: * linkerd/linkerd2#8733 * #3504 Signed-off-by: katelyn martin <kate@buoyant.io> X-Ref: tower-rs/tower#815 X-Ref: tower-rs/tower#817 X-Ref: tower-rs/tower#818 X-Ref: tower-rs/tower#819
Currently, the
DetectHttp
layer always builds both HTTP and TCPstacks, even though only one is needed. Furthermore, we lose the
information about which protocol was detected--it's never passed into
the http stack (and instead must be inferred from individual requests).
This change modifies the
DetectHttp
layer to only build inner servicesas needed, after the protocol has been detected.
Furthermore, outbound caching has been moved up from the tcp balancer to
wrap the HTTP detection layer. This was necessary so that the tcp
balancer could be built instantaneously as a NewService while still
being cloneable. This means that, for each outbound target ip:port, we
will reuse HTTP and TCP stacks.
However, HTTP routing remains dependent on per-request metadata. This is
expected to change in a followup.