fix: a refused caching opt-in re-tracks the connection and reruns the request - #22
Merged
Merged
Conversation
Probe::publish had one caller; conduct stores the four atomics itself.
… request Twice in ~40 matrix runs a backend answered a fill's CLIENT CACHING YES with the tracking-off error, and the client saw that error; ~120 runs of the same cell in isolation never reproduce it, and every fill path reads as covered by the dial-time frame or the rearm. The failure is now harmless: the pairing loop reports a refused opt-in (logged at warn with the connection), the connection sends its current tracking frame again on its own account (a pending entry with no sink), and the writer reruns the request without the opt-in when nothing later holds a sequence, abandoning the fill. Connections look the frame up from the shared map at handshake instead of carrying one captured at dial time.
…equest A fan-out part whose opt-in the connection refused reran nowhere: its error merged into the client's reply. The part now scatters again to its node without the head, after the fill it held is abandoned by the error reply. The single-request rerun keyed on the error text alone, so an application error of the same text on a plain command would run it twice; the rerun now requires the in-flight entry to hold a fill. The no-owner path leaked the detached fill and left the topology stale; it abandons the fill and asks for a refresh like the other bounces. The refusal log moves next to the tracking frame write, so it is seen only when the frame is on the wire, and pair_replies loses the address it carried for it.
The rerun indexed the fresh topology with the node index the part was planned with; a refresh between the plan and the refusal can renumber the nodes, sending the rerun elsewhere. The part's slot is resolved against the fresh topology instead, as the single-request rerun does.
Four opus readers walked every source file in full for clones, copies and borrow shapes; the rows on a per-command path that a borrow or a move serves as well are applied here, the rest are recorded. - a remembered script is stored as its SCRIPT LOAD frame, built once from the request, so a watched EVALSHA takes a refcount instead of a copy of the body, and SCRIPT LOAD copies the body once instead of twice - the dispatch subcommand lookup clones the positioned argument iterator instead of re-scanning the frame header and the name - a watched command checks its connection through the lease borrow instead of cloning the Rc to read one flag - the writer's ready buffer gets the excursion valve the parked ring, the timings and the fan-out gates already have IOV_STACK stays at 16: 97743c8 measured 64 slower by 0.2-3.0% on the default cells, since the array is initialized per writev and default batches average 1.3 frames. Also folds pair_replies' comment to one line.
The topology the rerun resolves its owner from can be one slot move stale, so the rerun could meet MOVED or ASK and merge it into the client's reply. A rerun's reply now takes the same one redirect a first-pass part takes; the redirect step is one helper for both.
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.
Three things: the last open row of the loc-justify cut-list, the reply cache's refused-opt-in intermittent, and an ownership walkthrough of the whole crate.
Probe::publishis inlined intoconduct, which already stores the preference (−3 lines).CLIENT CACHING YESwith the tracking-off error is now harmless: the pairing loop reports the refusal, the connection sends its current tracking frame again on its own account (logged at warn once the frame is written), and the request runs again without the opt-in, abandoning its fill, so the client sees its reply and nothing fills. A fan-out part reruns at its slot's current owner; a single request reruns while it is the session's last in flight, since a rerun behind later commands would reorder them. Only a request that held a fill reruns, so an application error of the same text never runs a command twice. Connections resolve the tracking frame from the shared map at handshake instead of carrying one captured at dial time.SCRIPT LOADframe, so a watched EVALSHA takes a refcount instead of copying the body; the dispatch subcommand lookup clones the positioned argument iterator instead of re-scanning the frame; a watched command checks its connection through the lease borrow instead of cloning theRc; the writer's ready buffer gets the capacity valve its siblings have;IOV_STACKgrows to 64 so a pipelined batch keeps its iovec on the stack.The refusal was seen twice in about forty compatibility-matrix runs of the redis 8.2 / reply-cache cell and never in about 120 isolated runs of the same cell; every fill path reads as covered by the dial-time frame or the tracker's rearm, so this closes the effect rather than a window that could be named. Nothing new runs on the reply path outside the existing error branches.
Measured on a 128-master rig, 64 workers, three arms rotated over nine rounds, P16 1000 conns: the fix against master GET 8.65M → 8.72M (+0.9%, A/A +0.05%), SET 9.01M → 9.04M (+0.4%, A/A −0.2%); the final head GET 8.67M → 8.77M (+1.3%, median +0.0%, A/A +0.9%), SET 9.10M → 9.08M (−0.2%, A/A −0.6%). Flat inside the A/A floor both times.
Gates: unit 118; the 15-cell compatibility matrix green; ct1 IT ×8 green, functional suite 165 / 163 passed. Review converged.