PrefixSet: git-backed CIDR files via gitshallow, refreshed every 47 min.
DomainSet: takes pre-parsed (staticPrefixes, domains []string); resolves
hostnames every 5 min, retaining stale IPs on failure. Both use
atomic-swapped ipcohort.Cohort for lock-free reads.
ParseDomainSet(RecordReader) parses rows into prefixes and hostnames;
caller constructs the record reader (e.g. csv.Reader with delimiter,
comment char). RecordReader interface is satisfied by *csv.Reader.
commaify uses plain string concat (integers max ~7 groups).
Resolves A, CNAME chain, and SRV records via miekg/dns. Queries all
configured servers in parallel and returns the first success. Falls
back to OpenDNS/Cloudflare/Quad9 when /etc/resolv.conf is absent or empty.
- Default depth changed from 1 to 3: enough overlap for delta
compression even if the remote pushes 2 commits between fetches
- Keep --depth on fetch (not just clone) so shallow stays windowed
- GCInterval default changed: 0 now means every fetch (was disabled)
- GC sequence: reflog expire → prune → gc --prune=now
Expire makes old shallow entries' objects unreachable, prune removes
them cheaply (5MB RSS), then gc delta-compresses what remains (~10MB)
- Only GC when HEAD actually changed (skip no-op fetches)
- Dropped GCAggressive (no benefit in shallow clones)
Periodically-updated shallow clone for repos used as a data source.
One Fetch verb (clone-or-pull-or-skip) drives both the Fetcher
interface and per-File handles. MaxAge gates on FETCH_HEAD mtime so
short-lived CLI invocations don't hammer the remote across process
restarts; an in-memory 1s debounce coalesces concurrent Fetches.
Hardens git invocations against argument injection: validates
branch/ref shape, rejects URLs starting with '-', and uses '--'
separators before positional args. Userinfo in URL is redacted from
error output (both args echo and git's own messages).
IPv4 host + CIDR membership testing backed by sorted /32 binary search
and a linear scan of CIDR ranges. CSV/text loader with bounded reads
(maxBytes is required so callers pick a limit appropriate to their
input source — 1 KB allowlists through 100+ MB threat feeds). Returns
errors for unparseable IPs so callers can decide fail-open vs
fail-closed; partial cohorts are returned alongside parse errors.
