Conversation
Tower's Buffer middleware is used to make a Service shareable across tasks. Within Linkerd, we've augmented the Buffer with failfast behavior, etc, as a Queue middleware. This queue plays a vital role in dispatching requests to load balancers, ensuring that load is shed when a balancer has no available endpoints. The tower Balancer additionally holds a Stream of endpoint updates that are processed as the balancer is driven to readiness. Crucially, this means that updates to the balancer can only occur while requests are being processed. We cannot eagerly drop defunct endpoints, nor can we eagerly connect to new endpoints. We have observed situations where long-lived, mostly-idle balancers can buffer discovery updates indefinitely, bloating memory and even forcing backpressure to the control plane. To correct this behavior, this change introduces the PoolQueue middleware. The PoolQueue is based on the tower's Buffer, but instead of composing over an inner Service, it composes over an inner Pool. Pool is a new interface that provides an additional interface to update the pool's members and to drive all pending endpoints in the pool to be ready (decoupling the semantics of Service::poll_ready and Pool::poll_pool). Pool implementations will typically hold a ReadyCache of inner endpoints (as the tower balancer does). This change, however, does not include the concrete implementation of a Pool to be replace the balancer. A p2c pool will be introduced in a followup change. This change has the added benefit of simplifying the endpoint discovery pipeline. We currently process Updates (including an array of endpoints) from the discovery API and convert that into a stream of discrete endpoint updates for the balancer, requiring redundant caching. The Pool interface processes Updates directly, so there is no need for the extra translation.
Following #2540, which introduces a new PoolQueue and Pool interface, this change introduces a P2cPool implementation that replaces Tower's p2c balancer (using the same underlying ReadyCache and p2c implementations). This balancer implementation is currently unused. It will be integrated in a follow-up change.
Codecov Report
Additional details and impacted files@@ Coverage Diff @@
## main #2541 +/- ##
==========================================
+ Coverage 67.35% 67.57% +0.22%
==========================================
Files 329 330 +1
Lines 14652 14764 +112
==========================================
+ Hits 9869 9977 +108
- Misses 4783 4787 +4
... and 10 files with indirect coverage changes Continue to review full report in Codecov by Sentry.
|
olix0r
added a commit
to linkerd/linkerd2
that referenced
this pull request
Dec 13, 2023
This change culminates recent work to restructure the balancer to use a PoolQueue so that balancer changes may occur independently of request processing. This replaces independent discovery buffering so that the balancer task is responsible for polling discovery streams without independent buffering. Requests are buffered and processed as soon as the pool has available backends. Fail-fast circuit breaking is enforced on the balancer's queue so that requests can't get stuck in a queue indefinitely. In general, the new balancer is instrumented directly with metrics, and the relevant metric name prefix and labelset is provided by the stack. In addition to detailed queue metrics including request (in-queue) latency histograms, but also failfast states, discovery updates counts, and balancer endpoint pool sizes. --- * outbound: Move queues into the concrete stack (linkerd/linkerd2-proxy#2539) * metrics: Remove unused features (linkerd/linkerd2-proxy#2542) * Add the PoolQueue middleware (linkerd/linkerd2-proxy#2540) * ci: Fixup codecov config (linkerd/linkerd2-proxy#2545) * ci: Cancel prior runs (linkerd/linkerd2-proxy#2546) * ci: Skip ARM builds during non-release CI (linkerd/linkerd2-proxy#2547) * deps: Update tokio, tonic, and prost (linkerd/linkerd2-proxy#2544) * build(deps): bump tj-actions/changed-files from 40.2.0 to 40.2.1 (linkerd/linkerd2-proxy#2549) * metrics: Use prometheus-client for proxy_build_info (linkerd/linkerd2-proxy#2551) * balance: Add a p2c Pool implementation (linkerd/linkerd2-proxy#2541) * metrics: Export process metrics using prometheus-client (linkerd/linkerd2-proxy#2552) * linkerd_identity: split `linkerd_identity::Id` into DNS and URI variants (linkerd/linkerd2-proxy#2538) * outbound: Move HTTP balancer into its own module (linkerd/linkerd2-proxy#2554) * app: Setup prom registry for use in balancers (linkerd/linkerd2-proxy#2555) * vscode: Move workspace settings to devcontainer (linkerd/linkerd2-proxy#2557) * build(deps): bump tj-actions/changed-files from 40.2.1 to 40.2.2 (linkerd/linkerd2-proxy#2556) * balance: Instrument metrics in pool balancer (linkerd/linkerd2-proxy#2558) * Enable PoolQueue balancer (linkerd/linkerd2-proxy#2559) Signed-off-by: Oliver Gould <[email protected]>
olix0r
added a commit
to linkerd/linkerd2
that referenced
this pull request
Dec 14, 2023
This change culminates recent work to restructure the balancer to use a PoolQueue so that balancer changes may occur independently of request processing. This replaces independent discovery buffering so that the balancer task is responsible for polling discovery streams without independent buffering. Requests are buffered and processed as soon as the pool has available backends. Fail-fast circuit breaking is enforced on the balancer's queue so that requests can't get stuck in a queue indefinitely. In general, the new balancer is instrumented directly with metrics, and the relevant metric name prefix and labelset is provided by the stack. In addition to detailed queue metrics including request (in-queue) latency histograms, but also failfast states, discovery updates counts, and balancer endpoint pool sizes. --- * outbound: Move queues into the concrete stack (linkerd/linkerd2-proxy#2539) * metrics: Remove unused features (linkerd/linkerd2-proxy#2542) * Add the PoolQueue middleware (linkerd/linkerd2-proxy#2540) * ci: Fixup codecov config (linkerd/linkerd2-proxy#2545) * ci: Cancel prior runs (linkerd/linkerd2-proxy#2546) * ci: Skip ARM builds during non-release CI (linkerd/linkerd2-proxy#2547) * deps: Update tokio, tonic, and prost (linkerd/linkerd2-proxy#2544) * build(deps): bump tj-actions/changed-files from 40.2.0 to 40.2.1 (linkerd/linkerd2-proxy#2549) * metrics: Use prometheus-client for proxy_build_info (linkerd/linkerd2-proxy#2551) * balance: Add a p2c Pool implementation (linkerd/linkerd2-proxy#2541) * metrics: Export process metrics using prometheus-client (linkerd/linkerd2-proxy#2552) * linkerd_identity: split `linkerd_identity::Id` into DNS and URI variants (linkerd/linkerd2-proxy#2538) * outbound: Move HTTP balancer into its own module (linkerd/linkerd2-proxy#2554) * app: Setup prom registry for use in balancers (linkerd/linkerd2-proxy#2555) * vscode: Move workspace settings to devcontainer (linkerd/linkerd2-proxy#2557) * build(deps): bump tj-actions/changed-files from 40.2.1 to 40.2.2 (linkerd/linkerd2-proxy#2556) * balance: Instrument metrics in pool balancer (linkerd/linkerd2-proxy#2558) * Enable PoolQueue balancer (linkerd/linkerd2-proxy#2559) Signed-off-by: Oliver Gould <[email protected]>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Following #2540, which introduces a new PoolQueue and Pool interface, this change introduces a P2cPool implementation that replaces Tower's p2c balancer (using the same underlying ReadyCache and p2c implementations).
This balancer implementation is currently unused. It will be integrated in a follow-up change.