mirror of
https://github.com/hanzoai/s3.git
synced 2026-10-30 14:59:09 +00:00
zooqueen
dd74d3e6d6
hanzoai/s3 was an AGPL-3.0 MinIO fork whose network-copyleft blocks embedding
and reselling as a managed service. Replace it wholesale with a fork of
SeaweedFS (github.com/seaweedfs/seaweedfs, Apache-2.0, Go) which ships a
built-in S3 gateway + Filer.
- License: AGPL-3.0 -> Apache-2.0 (upstream LICENSE kept verbatim, incl.
"Copyright 2025 Chris Lu"); NOTICE added crediting SeaweedFS. No AGPL/BUSL/
SSPL/GPL or enterprise ("ee") files remain.
- Module: go.mod path set to github.com/hanzoai/s3 with a self-replace
(replace github.com/seaweedfs/seaweedfs => ./) so the ~1800 internal Go files
keep their upstream import paths with zero churn. Internal identifiers and
import paths are intentionally NOT renamed.
- CLI rebrand: the binary is built as `s3` (was `weed`); root usage, per-command
help/usage/flag strings, shell autocomplete program name, and the product name
in CLI output now read "Hanzo S3" / `s3`. Users never see `weed`.
(e.g. `s3 server`, `s3 filer`, `s3 master`, `s3 volume`, `s3 s3`, `s3 mount`.)
- Build: `make` (or `go build -o s3 ./weed`) builds the `s3` binary green;
`cmd/weed-db` and `cmd/weed-sql` also build; `go vet ./weed/command/` clean.
- README/LLM.md rewritten for Hanzo S3; AGENTS.md/CLAUDE.md symlinks preserved.
- Dropped a 33MB prebuilt Mach-O `stress` binary upstream had committed; `.gitignore`
updated to ignore the `s3` build output.
67 lines
2.0 KiB
Go
67 lines
2.0 KiB
Go
package filer
|
|
|
|
import (
|
|
"sync/atomic"
|
|
)
|
|
|
|
type ReaderPattern struct {
|
|
isSequentialCounter int64
|
|
readFrontier int64 // highest (offset+size) observed across reads
|
|
}
|
|
|
|
const ModeChangeLimit = 3
|
|
|
|
// SeqTolerance: a read whose start is within this many bytes of the current read
|
|
// frontier still counts as sequential. Using a tolerance window rather than
|
|
// strict contiguity absorbs reordered/concurrent readahead (multiple ReadAt can
|
|
// be in flight at once) while still rejecting far random jumps.
|
|
const SeqTolerance = 8 << 20 // 8 MiB
|
|
|
|
// For streaming read: only cache the first chunk
|
|
// For random read: only fetch the requested range, instead of the whole chunk
|
|
|
|
func NewReaderPattern() *ReaderPattern {
|
|
return &ReaderPattern{
|
|
isSequentialCounter: 0,
|
|
readFrontier: 0,
|
|
}
|
|
}
|
|
|
|
func (rp *ReaderPattern) MonitorReadAt(offset int64, size int) {
|
|
// Advance the frontier to max(frontier, offset+size) and capture, in the same
|
|
// CAS loop, the pre-image this read is judged against. Reading the frontier
|
|
// inside the loop (rather than once up front) keeps `diff` below comparing
|
|
// against the freshest value even if a concurrent readahead advances the
|
|
// frontier while we loop. Lock-free, consistent with the rest of this type.
|
|
end := offset + int64(size)
|
|
var frontier int64
|
|
for {
|
|
frontier = atomic.LoadInt64(&rp.readFrontier)
|
|
if end <= frontier || atomic.CompareAndSwapInt64(&rp.readFrontier, frontier, end) {
|
|
break
|
|
}
|
|
}
|
|
|
|
// near = this read starts within SeqTolerance of where reads had reached.
|
|
// Hysteresis (the ±ModeChangeLimit counter) keeps a single outlier read from
|
|
// flipping the mode.
|
|
diff := offset - frontier
|
|
if diff < 0 {
|
|
diff = -diff
|
|
}
|
|
counter := atomic.LoadInt64(&rp.isSequentialCounter)
|
|
if diff <= SeqTolerance {
|
|
if counter < ModeChangeLimit {
|
|
atomic.AddInt64(&rp.isSequentialCounter, 1)
|
|
}
|
|
} else {
|
|
if counter > -ModeChangeLimit {
|
|
atomic.AddInt64(&rp.isSequentialCounter, -1)
|
|
}
|
|
}
|
|
}
|
|
|
|
func (rp *ReaderPattern) IsRandomMode() bool {
|
|
return atomic.LoadInt64(&rp.isSequentialCounter) < 0
|
|
}
|