mirror of
https://github.com/coredns/coredns.git
synced 2026-08-20 23:08:28 -04:00
50ffa1ecf9
remapStringRewriter matches a record name against orig and its sub domains. The
sub domain check was strings.HasSuffix(src, "."+r.orig), which built the
dot-prefixed string on every call and threw it away. Match the label boundary by
index instead: src is a sub domain of orig when orig sits at the end of src with
a "." immediately before it, which is exactly what the HasSuffix call tested.
Go concatenates short strings into a 32-byte stack buffer, so "."+orig only
reached the heap once orig passed 31 bytes. Below that the temporary was free
and this saves a few ns per record. Above it, 48 B was allocated per record.
Kubernetes service names are past the threshold -
my-service.my-namespace.svc.cluster.local. is 42 bytes - and those are the names
an auto rule rewrites to when a Corefile maps an external name onto an
in-cluster one. orig is the name the question was rewritten to, so whether a
deployment sees the allocation is a property of its Corefile, not its queries.
Caching "."+orig on the rewriter instead does not work: responseRuleFor
constructs a new rewriter for every request an auto name rule rewrites, so the
concatenation would run once per request rather than once per rule, and escapes
to the heap from there. That buys per-record work with a per-request allocation
and regresses every response short enough not to amortize it.
Per record, one rewriteString call on an existing rewriter:
name master this PR
RemapStringRewriter/short/match 186.6n 16 B/1 120.4n 16 B/1 -35%
RemapStringRewriter/short/nomatch 61.5n 0 B/0 16.5n 0 B/0 -73%
RemapStringRewriter/long/match 381.9n 72 B/2 165.3n 24 B/1 -57%
RemapStringRewriter/long/nomatch 219.1n 48 B/1 18.7n 0 B/0 -91%
Per request - rewrite the question, build the response rules, apply them to the
answer:
name master this PR
AutoNameRuleResponse/exact 935n 96 B/4 945n 96 B/4 ~
AutoNameRuleResponse/subdomain 1.248µ 112 B/5 1.242µ 112 B/5 ~
AutoNameRuleResponse/nomatch 1.016µ 96 B/4 945n 96 B/4 -7%
AutoNameRuleResponse/subdomain-8 3.376µ 224 B/12 2.891µ 224 B/12 -14%
AutoNameRuleResponse/k8s/subdomain 1.611µ 176 B/6 1.380µ 128 B/5 -14%
AutoNameRuleResponse/k8s/subdomain-8 5.144µ 672 B/20 3.305µ 288 B/12 -36%
benchstat over 8 runs, i7-1065G7. With short names this is flat at the request
level: one rewriteString call is small next to the four allocations that
building the rules costs. The saving is per record and per byte of name, so it
shows up where responses carry several records and the rewritten-to name is
long.
This applies to exact, prefix, substring and regex name rules with answer auto.
suffix rules build a suffixStringRewriter instead and are not affected.
TestRemapStringRewriter pins the label boundary semantics the index arithmetic
now carries, notably that notexample.com. is not a sub domain of example.com.
It passes against the previous implementation too.
Signed-off-by: Manuel Rüger <manuel@rueg.eu>