Files
seaweedfs/weed/util/log_buffer/log_buffer.go
T
5d5fcdf07b fix(filer): bound aggregated metadata reads by peer watermarks (#10803)
* fix(filer): watermark-bound aggregated metadata subscription against multi-source merge races

The aggregated metadata subscription (SubscribeMetadata) merges per-filer
sources that become readable at independent paces, but tracks its progress
with a single scalar cursor. Once the cursor passes a timestamp T, anything
a source materializes below T afterwards is silently skipped: a peer
recovering from a stall re-inserts its backlog late (late ring merge), and
a source's flush can land a log file, or a later chunk of the same file,
after a subscriber's disk pass listed the files (late persisted-log
landing). This is the residual documented in #10501.

Bound the subscriber's two read paths by what every source has provably
made visible, each with its own watermark:

- Delivery low-watermark -> in-memory reads. The meta aggregator tracks,
  per subscribed peer (self included), the newest timestamp received on
  that peer's stream - real events, or idle heartbeats (peer streams now
  opt into ClientSupportsIdleHeartbeat). The aggregated ring is complete
  up to the minimum across peers; in-memory reads hold at it.
- Flush low-watermark -> persisted-log reads. Each filer reports its local
  log-buffer flush watermark on its stream: a new flushed_ts_ns response
  field, carried on idle heartbeats and on periodic flush reports (gated
  on ClientSupportsIdleHeartbeat). Disk passes freeze the minimum across
  peers before listing the log files and hold at it; the day-boundary
  cursor jump and the metadata-chunks ref listing are bounded the same
  way, the latter at minute-file granularity.
- Held reads keep the cursor at the last entry actually delivered and
  retry; the retry re-lists the log files, which is what picks up a
  late-landing file. Both watermarks are relaxed by the settled horizon
  (2 x LogFlushInterval) as a liveness escape, so a peer stalled beyond it
  delays subscribers by at most the horizon instead of forever - any loss
  that escape allows was unconditional before.

With reads held at the flush watermark, a disk advance below it is proven
complete on every peer's disk, so the unproven-crossing counter now only
counts crossings the horizon escape allowed past a stalled peer.

Live delivery on the aggregated stream may lag by up to the idle-heartbeat
interval when some peers are quiet; SubscribeLocalMetadata consumers are
unaffected.

* fix(filer): resume evicted aggregated readers from an original-space disk anchor

The aggregated ring rewrites out-of-order peer arrivals to its head, so a
subscriber tailing it advances its cursor in bumped (arrival) timestamps,
while persisted logs keep original timestamps. When a slow reader's unread
window is evicted (e.g. a peer backlog flooding in after a stall) and the
reader falls back to disk, resuming from the bumped cursor skips every
original-space entry below it that memory never delivered - reproduced as
a ~66% silent loss on a 3-filer cluster with one peer's stream frozen for
~70s while the subscriber lagged.

Track a disk anchor: the newest original-space position the stream is
proven complete through. Disk passes advance it directly; contiguous
memory reads advance it to the peers' delivery low-watermark observed
before the read (per-peer streams are ordered, so everything with an
original timestamp at or below that watermark had already arrived and was
delivered). A reader kicked off the ring resumes the disk pass from the
anchor instead of the bumped cursor - redelivering what memory already
sent is within the subscription's at-least-once contract, skipping what
it never sent is not.

* fix(filer): close review findings on the peer-watermark subscription bounds

Four correctness holes found in review, one generated-file cleanup:

- The flush-through claim could assert durability for events still on
  their way into the buffer: an event is timestamped before notification
  work that can block, and only then appended. Track stamped-but-unappended
  events on the Filer (the stamp shares a lock with the reader, and appends
  are bumped monotonically past the buffer head), and cap the reported
  flush watermark just below the oldest in-flight stamp.

- Removing a peer deleted its watermark entries while its stream kept
  running: its next signal recreated the deleted entry, which then pinned
  the low-watermark forever once the stream died. Watermarks now advance
  only for tracked peers, and peer removal cancels the subscription
  context so the stream stops feeding the aggregated buffer promptly.

- The pipelined sender folded flush reports (TsNs 0 reads as far behind)
  into batch Events tails, where the aggregator's nil-notification guard
  dropped them - a busy backlog replay could starve the flush watermark
  until the settled-horizon escape opened a loss window. Control messages
  are now unbatchable on the sender, and the receiver also reads watermark
  state off nested batch entries as belt and braces.

- A give-up skip's cursor was not anchored, so the next eviction rewind
  undid the counted decision and re-entered the same park forever when the
  evicted window carried bumped timestamps. The anchor now follows give-up
  skips; an anchored cursor makes the rewind a no-op and keeps the gap
  machinery's re-arm onto the retained window reachable.

- Regenerated-file churn from a different protoc-gen-go-vtproto version is
  dropped: the vtproto file is upstream's, plus only the flushed_ts_ns
  marshal/size/unmarshal cases in the same generator style.

New tests pin the in-flight floor, the no-resurrection rule for removed
peers, and that control messages are never nested in batches.

* fix(filer): keep a removed peer's watermarks through a grace period

Deleting a peer's watermark entries the moment the master removes it
reopened the loss the watermarks exist to prevent: a filer frozen or
partitioned long enough to miss master heartbeats is removed from the
cluster, its unflushed events still exist, and with its entries gone the
low-watermarks snap forward to the healthy peers - subscribers advance
past the absent peer's window and its late-landing log files are silently
skipped. Reproduced on a 3-filer cluster: freezing two filers for ~70s got
them removed ~28s in, and a catching-up subscriber lost their entire
overlapping window.

Removal now only marks the peer; its watermarks keep participating in the
low-watermarks for a grace period (2 x LogFlushInterval, matching the
subscribe loops' settled horizon, which already bounds a stale watermark's
influence meanwhile). A re-added peer clears the mark and continues its
values monotonically - the flap case costs nothing. A peer that stays gone
is dropped when the grace expires, so a decommission cannot pin the
low-watermarks, and a dropped peer's straggling signals cannot resurrect
its entry.

* fix(filer): cap delivery heartbeats by the in-flight floor; harden stamps

Second review pass on the watermark bounds:

- Idle heartbeats on the local stream claimed delivery-completeness
  through "now" while an event could still sit stamped-but-unappended
  behind blocking notification work. A peer aggregator turns that claim
  into its delivery low-watermark, so it could advance (and anchor
  credits with it) past an event that had not been streamed yet. The
  heartbeat timestamp is now capped just below the oldest in-flight
  stamp, like the flush claim already was.

- In-flight stamps are forced monotonic against the registry's own
  history, so a wall-clock step backwards cannot slip a new stamp under
  an already-sampled floor. The cross-goroutine ordering still shares
  the meta log's global forward-clock assumption; the comments now say
  so instead of overclaiming.

- Duplicate removal notifications no longer refresh a removed peer's
  grace deadline: the first removal time wins, so a decommissioned peer
  cannot sit in the watermark sets forever on repeated updates.

- A failed buffer append clears the event's in-flight stamp on purpose:
  the event is dropped from the change stream entirely (a pre-existing
  defect of the append path, loudly logged), and a watermark waiting for
  it would pin this filer's claims forever. The comments now state the
  decision instead of implying the failure cannot happen.

* docs(filer): tighten the watermark comments

Comment-only: compress the narrative comments added on this branch down
to their load-bearing invariants, and fix one stale sentence (peer
removal no longer deletes the watermark entries immediately). No code
changes.

* fix(filer): subscribe to the local filer before remote peers

Self's events reach the aggregated buffer only through the aggregator's
own subscription to it, but bootstrap only seeded the peers the master
already listed - and self's master registration races that listing, so
the watermark set could hold remote peers without self. Once the remotes
signalled, the low-watermarks would claim completeness for a stream that
was still missing a merge source, letting aggregated subscribers advance
past the local filer's events before its subscription started.

Seed self first, unconditionally: before that the watermark set is empty
(a documented safe state - reads hold at the settled horizon), and after
it the set can never be remotes-only. The later master update for self,
or a duplicate in the listed peers, is a no-op via the already-followed
check in OnPeerUpdate.

* fix(filer): fence watermark claims against wall-clock regression

Record issued heartbeat/flush claims in the in-flight registry and stamp
later events above them, so a backward clock step cannot land an event
under a watermark a peer has already advanced to.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(filer): re-check the buffer head after fencing heartbeat claims

An event appended between the caught-up check and the delivery claim
was covered by the claim but not yet sent on the stream. The claims
fence later stamps, so re-checking the head after them proves every
covered event was already sent before the heartbeat.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(filer): cross the aggregated ring's pre-subscription range only on proof

The eviction gate and the gap proofs read "nothing evicted yet" as "memory
holds everything after the cursor". That is false for the merge-fed
aggregated ring, which is born empty while every peer's history sits on
disk: before the ring's first real eviction, a subscriber whose cursor was
still below the bounded chunk pass's listing stop was served the ring's
earliest entry inclusively, silently skipping the withheld pre-restart
files - and the idle-wait callback credited the delivery low-watermark to
the disk anchor in the same disconnected state.

Mark everything at or below the subscriptions' start as evicted when the
aggregator is built, credit the anchor only once the run is connected to
the ring, and give the aggregated gap pass a real proof to cross the
marked boundary with: each disk pass's proven coverage (the peer flush
low-watermark capped by the pass's listing bound). An empty pass whose
proof reaches the eviction watermark crosses to it silently - no park, no
loss counter - so the mark costs a bounded catch-up delay instead of the
15-minute give-up.

* fix(filer): keep shipped chunk tails at or below the hold point

A log file spans past its named minute (window start plus up to a flush
interval), and chunk-mode clients apply a shipped file whole - so a file
tail past the hold point can become a persisted client checkpoint beyond
what every peer has proven, and a crash inside that window resumes past
another peer's late-but-in-contract flush. Stop the ref listing a minute
plus a flush interval below the hold; the withheld band is served by the
memory pass (ring retention far exceeds it) or by later passes as the
hold advances, so freshness is unchanged. A frozen peer flushing one
window that spans its whole freeze can still overshoot; that residual is
bounded by the freeze and needs a crash inside it.

* docs(filer): trim the review-fix comments

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Chris Lu <chris.lu@gmail.com>
2026-08-19 18:38:46 -07:00

1434 lines
51 KiB
Go

package log_buffer
import (
"bytes"
"encoding/binary"
"fmt"
"math"
"sync"
"sync/atomic"
"time"
"github.com/seaweedfs/seaweedfs/weed/glog"
"github.com/seaweedfs/seaweedfs/weed/pb/filer_pb"
"github.com/seaweedfs/seaweedfs/weed/pb/mq_pb"
"github.com/seaweedfs/seaweedfs/weed/util"
"github.com/seaweedfs/seaweedfs/weed/util/mem"
)
const BufferSize = 8 * 1024 * 1024
const PreviousBufferCount = 4
// EvictionGatedOffset is a sentinel cursor offset (-2..-6 are taken by other
// sentinels) that reads like the plain -2 sentinel except below the eviction
// watermark: there ReadFromBuffer refuses with ResumeFromDiskError instead of
// silently serving from the earliest retained window. The check runs under the
// read lock, atomically with the serve decision, which callers cannot do from
// outside - an eviction can land between any caller-side check and the read.
const EvictionGatedOffset = -7
// flushQueueDepth bounds queued flush copies (BufferSize each); a full queue
// blocks producers, so a stalled flush backpressures writers instead of
// pinning hundreds of buffer copies.
const flushQueueDepth = 16
// flushQueueBudget bounds the same queue in bytes. Counting copies only holds
// if every copy is a window's worth: an entry larger than BufferSize grows its
// window to fit, and a queue of those multiplies straight through — sixteen
// 100 MB windows is 1.6 GB of flush copies alone. The ceiling is the one the
// depth was chosen for, so ordinary windows still queue sixteen deep.
//
// This covers the flush queue only. A sealed window stays reachable through
// prevBuffers for PreviousBufferCount more seals, and readers may take a
// snapshot of it, so an oversized entry still costs several times its size
// before it falls out of the ring.
const flushQueueBudget = flushQueueDepth * BufferSize
// Errors that can be returned by log buffer operations
var (
// ErrBufferCorrupted indicates the log buffer contains corrupted data
ErrBufferCorrupted = fmt.Errorf("log buffer is corrupted")
)
type dataToFlush struct {
startTime time.Time
stopTime time.Time
data []byte // slab from mem.Allocate; returned via mem.Free after flush
minOffset int64
maxOffset int64
seq uint64 // seal order, so the budget admits windows in order
budget int // bytes reserved from flushBudget, released after the flush
done chan struct{} // Signal when flush completes
}
// flushBudget accounts the bytes of sealed windows waiting to be written, so a
// producer waits for the queue to drain instead of adding another copy to it.
// Windows are admitted strictly in seal order: a producer parked here can wait
// seconds, and letting a later window overtake an earlier one would hand
// loopFlush the windows out of order.
type flushBudget struct {
mu sync.Mutex
cond *sync.Cond
limit int
queued int
nextSeq uint64
closed bool
}
func newFlushBudget(limit int) *flushBudget {
b := &flushBudget{limit: limit}
b.cond = sync.NewCond(&b.mu)
return b
}
// reserve blocks until it is this window's turn and its bytes fit under the
// limit, then returns the amount to hand back to release. A window larger than
// the whole budget is admitted on its own once the queue empties, so an
// oversized entry still gets through.
func (b *flushBudget) reserve(seq uint64, n int) int {
if n > b.limit {
n = b.limit
}
b.mu.Lock()
defer b.mu.Unlock()
for !b.closed && (seq != b.nextSeq || (b.queued > 0 && b.queued+n > b.limit)) {
b.cond.Wait()
}
if seq == b.nextSeq {
b.nextSeq++
}
b.queued += n
b.cond.Broadcast() // wake whoever is next in line
return n
}
// waitForRoom parks until the queue has headroom for a window of n bytes,
// without charging anything. A window is copied into its slab while the write
// lock is held, before queueFlush gets to reserve, so a burst of concurrent
// oversized writers would each be holding a full copy in hand by the time they
// queue up -- memory the budget never sees. Large writers wait here first so
// they arrive at the seal a few at a time. This throttles the burst rather than
// bounding it: a writer that passes the check still seals unconditionally.
func (b *flushBudget) waitForRoom(n int) {
if n > b.limit {
n = b.limit
}
b.mu.Lock()
defer b.mu.Unlock()
for !b.closed && b.queued > 0 && b.queued+n > b.limit {
b.cond.Wait()
}
}
func (b *flushBudget) release(n int) {
if n == 0 {
return
}
b.mu.Lock()
b.queued -= n
b.mu.Unlock()
b.cond.Broadcast()
}
// close stops the budget from parking anyone, so shutdown is never held up by
// a producer waiting on a flush that will not run.
func (b *flushBudget) close() {
b.mu.Lock()
b.closed = true
b.mu.Unlock()
b.cond.Broadcast()
}
type EachLogEntryFuncType func(logEntry *filer_pb.LogEntry) (isDone bool, err error)
type EachLogEntryWithOffsetFuncType func(logEntry *filer_pb.LogEntry, offset int64) (isDone bool, err error)
type LogFlushFuncType func(logBuffer *LogBuffer, startTime, stopTime time.Time, buf []byte, minOffset, maxOffset int64)
type LogReadFromDiskFuncType func(startPosition MessagePosition, stopTsNs int64, eachLogEntryFn EachLogEntryFuncType) (lastReadPosition MessagePosition, isDone bool, err error)
// DiskChunkCache caches chunks of historical data read from disk
type DiskChunkCache struct {
mu sync.RWMutex
chunks map[int64]*CachedDiskChunk // Key: chunk start offset (aligned to chunkSize)
maxChunks int // Maximum number of chunks to cache
}
// CachedDiskChunk represents a cached chunk of disk data
type CachedDiskChunk struct {
startOffset int64
endOffset int64
messages []*filer_pb.LogEntry
lastAccess time.Time
}
type LogBuffer struct {
// 8-byte aligned fields
LastTsNs atomic.Int64
lastFlushTsNs atomic.Int64
lastFlushedOffset atomic.Int64 // Highest offset that has been flushed to disk (-1 = nothing flushed yet)
lastEvictedTsNs atomic.Int64 // Latest stopTime evicted from the sealed ring (0 = nothing evicted yet)
// lastEvictedTsNs in pre-bump timestamps: gap proofs compare disk cursors,
// which never see the bumped values out-of-order arrivals get.
lastEvictedOriginalTsNs atomic.Int64
curWindowMaxOriginalTsNs int64 // max pre-bump ts in the open window, under the write lock
offset int64
bufferStartOffset int64
minOffset int64
maxOffset int64
flushInterval time.Duration
startTime time.Time
stopTime time.Time
// Other fields
name string
prevBuffers *SealedBuffers
buf []byte
idx []int
pos int
sizeBuf []byte
flushFn LogFlushFuncType
ReadFromDiskFn LogReadFromDiskFuncType
notifyFn func()
// Per-subscriber notification channels for instant wake-up
subscribersMu sync.RWMutex
subscribers map[string]chan struct{} // subscriberID -> notification channel
// Notified only when a flush lands, for readers that cannot act on an append
flushSubscribers map[string]chan struct{}
isStopping *atomic.Bool
shutdownCh chan struct{} // closed by ShutdownLogBuffer to wake blocked subscribers
loopsDone sync.WaitGroup // loopFlush and loopInterval signal exit
isAllFlushed bool
flushChan chan *dataToFlush
flushBudget *flushBudget
flushSeq uint64 // seal counter, assigned under the write lock
// Offset range tracking for Kafka integration
hasOffsets bool
// Disk chunk cache for historical data reads
diskChunkCache *DiskChunkCache
// curSnap is a GC-owned copy of the current window's append-only prefix
// buf[:len(curSnap)], shared by all readers so each byte is copied once per
// window instead of once per reader. Extended lazily under curSnapMu; reset
// at seal. Existing holders keep their prefix slices, which never mutate.
curSnapMu sync.Mutex
curSnap []byte
sync.RWMutex
}
func NewLogBuffer(name string, flushInterval time.Duration, flushFn LogFlushFuncType,
readFromDiskFn LogReadFromDiskFuncType, notifyFn func()) *LogBuffer {
lb := &LogBuffer{
name: name,
prevBuffers: newSealedBuffers(PreviousBufferCount),
buf: make([]byte, BufferSize),
sizeBuf: make([]byte, 4),
flushInterval: flushInterval,
flushFn: flushFn,
ReadFromDiskFn: readFromDiskFn,
notifyFn: notifyFn,
subscribers: make(map[string]chan struct{}),
flushSubscribers: make(map[string]chan struct{}),
flushChan: make(chan *dataToFlush, flushQueueDepth),
isStopping: new(atomic.Bool),
shutdownCh: make(chan struct{}),
offset: 0, // Will be initialized from existing data if available
flushBudget: newFlushBudget(flushQueueBudget),
diskChunkCache: &DiskChunkCache{
chunks: make(map[int64]*CachedDiskChunk),
maxChunks: 16, // Cache up to 16 chunks (configurable)
},
}
lb.lastFlushedOffset.Store(-1) // Nothing flushed to disk yet
lb.loopsDone.Add(2)
go lb.loopFlush()
go lb.loopInterval()
return lb
}
// RegisterSubscriber registers a subscriber for instant notifications when data is written
// Returns a channel that will receive notifications (<1ms latency)
func (logBuffer *LogBuffer) RegisterSubscriber(subscriberID string) chan struct{} {
logBuffer.subscribersMu.Lock()
defer logBuffer.subscribersMu.Unlock()
// Check if already registered
if existingChan, exists := logBuffer.subscribers[subscriberID]; exists {
return existingChan
}
// Create buffered channel (size 1) so notifications never block
notifyChan := make(chan struct{}, 1)
logBuffer.subscribers[subscriberID] = notifyChan
return notifyChan
}
// UnregisterSubscriber removes a subscriber and closes its notification channel
func (logBuffer *LogBuffer) UnregisterSubscriber(subscriberID string) {
logBuffer.subscribersMu.Lock()
defer logBuffer.subscribersMu.Unlock()
if ch, exists := logBuffer.subscribers[subscriberID]; exists {
close(ch)
delete(logBuffer.subscribers, subscriberID)
}
}
// RegisterFlushSubscriber registers a subscriber woken only when a flush lands.
// A reader waiting for data it can only get from disk has nothing to do with an
// append, and taking those wake-ups off the shared channel would cost it one
// scheduling round-trip per write - and keep that channel drained, so every
// writer's non-blocking send succeeds instead of falling through.
func (logBuffer *LogBuffer) RegisterFlushSubscriber(subscriberID string) chan struct{} {
logBuffer.subscribersMu.Lock()
defer logBuffer.subscribersMu.Unlock()
if existingChan, exists := logBuffer.flushSubscribers[subscriberID]; exists {
return existingChan
}
notifyChan := make(chan struct{}, 1)
logBuffer.flushSubscribers[subscriberID] = notifyChan
return notifyChan
}
// UnregisterFlushSubscriber removes a flush subscriber and closes its channel
func (logBuffer *LogBuffer) UnregisterFlushSubscriber(subscriberID string) {
logBuffer.subscribersMu.Lock()
defer logBuffer.subscribersMu.Unlock()
if ch, exists := logBuffer.flushSubscribers[subscriberID]; exists {
close(ch)
delete(logBuffer.flushSubscribers, subscriberID)
}
}
// IsOffsetInMemory checks if the given offset is available in the in-memory buffer
// Returns true if:
// 1. Offset is newer than what's been flushed to disk (must be in memory)
// 2. Offset is in current buffer or previous buffers (may be flushed but still in memory)
// Returns false if offset is older than memory buffers (only on disk)
func (logBuffer *LogBuffer) IsOffsetInMemory(offset int64) bool {
logBuffer.RLock()
defer logBuffer.RUnlock()
// Check if we're tracking offsets at all
if !logBuffer.hasOffsets {
return false // No offsets tracked yet
}
// OPTIMIZATION: If offset is newer than what's been flushed to disk,
// it MUST be in memory (not written to disk yet)
lastFlushed := logBuffer.lastFlushedOffset.Load()
if lastFlushed >= 0 && offset > lastFlushed {
return true
}
// Check if offset is in current buffer range AND buffer has data
// (data can be both on disk AND in memory during flush window)
if offset >= logBuffer.bufferStartOffset && offset <= logBuffer.offset {
// CRITICAL: Check if buffer actually has data (pos > 0)
// After flush, pos=0 but range is still valid - data is on disk, not in memory
if logBuffer.pos > 0 {
return true
}
// Buffer is empty (just flushed) - data is on disk
return false
}
// Check if offset is in previous buffers AND they have data
for _, buf := range logBuffer.prevBuffers.buffers {
if offset >= buf.startOffset && offset <= buf.offset {
// Check if prevBuffer actually has data
if buf.size > 0 {
return true
}
// Buffer is empty (flushed) - data is on disk
return false
}
}
// Offset is older than memory buffers - only available on disk
return false
}
// notifySubscribers sends notifications to all registered subscribers
// Non-blocking: uses select with default to avoid blocking on full channels
func (logBuffer *LogBuffer) notifySubscribers() {
logBuffer.subscribersMu.RLock()
defer logBuffer.subscribersMu.RUnlock()
if len(logBuffer.subscribers) == 0 {
return // No subscribers, skip notification
}
for _, notifyChan := range logBuffer.subscribers {
select {
case notifyChan <- struct{}{}:
// Notification sent successfully
default:
// Channel full - subscriber hasn't consumed previous notification yet
// This is OK because one notification is sufficient to wake the subscriber
}
}
}
// notifyFlushSubscribers wakes the readers that only care about a flush landing
func (logBuffer *LogBuffer) notifyFlushSubscribers() {
logBuffer.subscribersMu.RLock()
defer logBuffer.subscribersMu.RUnlock()
for _, notifyChan := range logBuffer.flushSubscribers {
select {
case notifyChan <- struct{}{}:
default:
}
}
}
// InitializeOffsetFromExistingData initializes the offset counter from existing data on disk
// This should be called after LogBuffer creation to ensure offset continuity on restart
func (logBuffer *LogBuffer) InitializeOffsetFromExistingData(getHighestOffsetFn func() (int64, error)) error {
if getHighestOffsetFn == nil {
return nil // No initialization function provided
}
highestOffset, err := getHighestOffsetFn()
if err != nil {
return nil // Continue with offset 0 if we can't read existing data
}
if highestOffset >= 0 {
// Set the next offset to be one after the highest existing offset
nextOffset := highestOffset + 1
logBuffer.offset = nextOffset
// bufferStartOffset should match offset after initialization
// This ensures that reads for old offsets (0...highestOffset) will trigger disk reads
// New data written after this will start at nextOffset
logBuffer.bufferStartOffset = nextOffset
// CRITICAL: Track that data [0...highestOffset] is on disk
logBuffer.lastFlushedOffset.Store(highestOffset)
// Set lastFlushTsNs to current time (we know data up to highestOffset is on disk)
logBuffer.lastFlushTsNs.Store(time.Now().UnixNano())
} else {
logBuffer.bufferStartOffset = 0 // Start from offset 0
// No data on disk yet
}
return nil
}
func (logBuffer *LogBuffer) AddToBuffer(message *mq_pb.DataMessage) error {
return logBuffer.AddDataToBuffer(message.Key, message.Value, message.TsNs)
}
// AddLogEntryToBuffer directly adds a LogEntry to the buffer, preserving offset information
func (logBuffer *LogBuffer) AddLogEntryToBuffer(logEntry *filer_pb.LogEntry) error {
if len(logEntry.Data) > BufferSize {
logBuffer.flushBudget.waitForRoom(len(logEntry.Data))
}
var toFlush *dataToFlush
var marshalErr error
logBuffer.Lock()
defer func() {
logBuffer.Unlock()
if toFlush != nil {
logBuffer.queueFlush(toFlush)
}
// Only notify if there was no error
if marshalErr == nil {
if logBuffer.notifyFn != nil {
logBuffer.notifyFn()
}
// Notify all registered subscribers instantly (<1ms latency)
logBuffer.notifySubscribers()
}
}()
processingTsNs := logEntry.TsNs
ts := time.Unix(0, processingTsNs)
originalTsNs := processingTsNs
// Handle timestamp collision inside lock (rare case)
if logBuffer.LastTsNs.Load() >= processingTsNs {
processingTsNs = logBuffer.LastTsNs.Add(1)
ts = time.Unix(0, processingTsNs)
// Re-marshal with corrected timestamp
logEntry.TsNs = processingTsNs
} else {
logBuffer.LastTsNs.Store(processingTsNs)
}
// Size is computed without allocating; the entry is marshaled straight into
// the buffer below via MarshalToSizedBufferVT.
size := logEntry.SizeVT()
if logBuffer.pos == 0 {
logBuffer.startTime = ts
// Reset offset tracking for new buffer
logBuffer.hasOffsets = false
}
// Track offset ranges for Kafka integration
// Use >= 0 to include offset 0 (first message in a topic)
if logEntry.Offset >= 0 {
if !logBuffer.hasOffsets {
logBuffer.minOffset = logEntry.Offset
logBuffer.maxOffset = logEntry.Offset
logBuffer.hasOffsets = true
} else {
if logEntry.Offset < logBuffer.minOffset {
logBuffer.minOffset = logEntry.Offset
}
if logEntry.Offset > logBuffer.maxOffset {
logBuffer.maxOffset = logEntry.Offset
}
}
}
if logBuffer.startTime.Add(logBuffer.flushInterval).Before(ts) || len(logBuffer.buf)-logBuffer.pos < size+4 {
toFlush = logBuffer.copyToFlush()
logBuffer.startTime = ts
if len(logBuffer.buf) < size+4 {
// Validate size to prevent integer overflow in computation BEFORE allocation
const maxBufferSize = 1 << 30 // 1 GiB practical limit
// The window is sized size+4, so that is what has to stay in bounds
if size < 0 || size > math.MaxInt-4 || size > maxBufferSize-4 {
marshalErr = fmt.Errorf("message size %d exceeds maximum allowed size", size)
glog.Errorf("%v", marshalErr)
return marshalErr
}
// Fit the entry exactly. Doubling left room for a second oversized
// record in the same window, which only doubles the flush copy and
// the snapshot taken of it.
logBuffer.buf = make([]byte, size+4)
}
}
logBuffer.stopTime = ts
// Marshal directly into the buffer, avoiding an intermediate slice and copy.
// On the (practically impossible) error the entry is dropped before idx/pos
// are advanced, leaving the buffer consistent.
if _, err := logEntry.MarshalToSizedBufferVT(logBuffer.buf[logBuffer.pos+4 : logBuffer.pos+4+size]); err != nil {
marshalErr = fmt.Errorf("failed to marshal LogEntry: %w", err)
glog.Errorf("%v", marshalErr)
return marshalErr
}
logBuffer.idx = append(logBuffer.idx, logBuffer.pos)
util.Uint32toBytes(logBuffer.sizeBuf, uint32(size))
copy(logBuffer.buf[logBuffer.pos:logBuffer.pos+4], logBuffer.sizeBuf)
logBuffer.pos += size + 4
// Only now is the entry's window known: a rollover above seals the previous
// window first, and crediting this timestamp before that hands it to the
// sealed window and loses it from the new one - corrupting the received-ts
// eviction watermark in both directions.
if originalTsNs > logBuffer.curWindowMaxOriginalTsNs {
logBuffer.curWindowMaxOriginalTsNs = originalTsNs
}
logBuffer.offset++
return nil
}
func (logBuffer *LogBuffer) AddDataToBuffer(partitionKey, data []byte, processingTsNs int64) error {
// An entry this large gets a window to itself, so it will seal and copy one;
// wait for the queue to have room before joining the queue for the lock.
if len(data) > BufferSize {
logBuffer.flushBudget.waitForRoom(len(data))
}
// PERFORMANCE OPTIMIZATION: Pre-process expensive operations OUTSIDE the lock
var ts time.Time
if processingTsNs == 0 {
ts = time.Now()
processingTsNs = ts.UnixNano()
} else {
ts = time.Unix(0, processingTsNs)
}
logEntry := &filer_pb.LogEntry{
TsNs: processingTsNs, // Will be updated if needed
PartitionKeyHash: util.HashToInt32(partitionKey),
Data: data,
Key: partitionKey,
}
var toFlush *dataToFlush
var marshalErr error
logBuffer.Lock()
defer func() {
logBuffer.Unlock()
if toFlush != nil {
logBuffer.queueFlush(toFlush)
}
// Only notify if there was no error
if marshalErr == nil {
if logBuffer.notifyFn != nil {
logBuffer.notifyFn()
}
// Notify all registered subscribers instantly (<1ms latency)
logBuffer.notifySubscribers()
}
}()
originalTsNs := processingTsNs
// Handle timestamp collision inside lock (rare case)
if logBuffer.LastTsNs.Load() >= processingTsNs {
processingTsNs = logBuffer.LastTsNs.Add(1)
ts = time.Unix(0, processingTsNs)
logEntry.TsNs = processingTsNs
} else {
logBuffer.LastTsNs.Store(processingTsNs)
}
// Set the offset in the LogEntry before marshaling
// This ensures the flushed data contains the correct offset information
// Note: This also enables AddToBuffer to work correctly with Kafka-style offset-based reads
logEntry.Offset = logBuffer.offset
// Size is computed without allocating; the entry is marshaled straight into
// the buffer below via MarshalToSizedBufferVT.
size := logEntry.SizeVT()
if logBuffer.pos == 0 {
logBuffer.startTime = ts
// Reset offset tracking for new buffer
logBuffer.hasOffsets = false
}
// Track offset ranges for Kafka integration
// Track the current offset being written
if !logBuffer.hasOffsets {
logBuffer.minOffset = logBuffer.offset
logBuffer.maxOffset = logBuffer.offset
logBuffer.hasOffsets = true
} else {
if logBuffer.offset < logBuffer.minOffset {
logBuffer.minOffset = logBuffer.offset
}
if logBuffer.offset > logBuffer.maxOffset {
logBuffer.maxOffset = logBuffer.offset
}
}
if logBuffer.startTime.Add(logBuffer.flushInterval).Before(ts) || len(logBuffer.buf)-logBuffer.pos < size+4 {
toFlush = logBuffer.copyToFlush()
logBuffer.startTime = ts
if len(logBuffer.buf) < size+4 {
// Validate size to prevent integer overflow in computation BEFORE allocation
const maxBufferSize = 1 << 30 // 1 GiB practical limit
// The window is sized size+4, so that is what has to stay in bounds
if size < 0 || size > math.MaxInt-4 || size > maxBufferSize-4 {
marshalErr = fmt.Errorf("message size %d exceeds maximum allowed size", size)
glog.Errorf("%v", marshalErr)
return marshalErr
}
// Fit the entry exactly. Doubling left room for a second oversized
// record in the same window, which only doubles the flush copy and
// the snapshot taken of it.
logBuffer.buf = make([]byte, size+4)
}
}
logBuffer.stopTime = ts
// Marshal directly into the buffer, avoiding an intermediate slice and copy.
// On the (practically impossible) error the entry is dropped before idx/pos
// are advanced, leaving the buffer consistent.
if _, err := logEntry.MarshalToSizedBufferVT(logBuffer.buf[logBuffer.pos+4 : logBuffer.pos+4+size]); err != nil {
marshalErr = fmt.Errorf("failed to marshal LogEntry: %w", err)
glog.Errorf("%v", marshalErr)
return marshalErr
}
logBuffer.idx = append(logBuffer.idx, logBuffer.pos)
util.Uint32toBytes(logBuffer.sizeBuf, uint32(size))
copy(logBuffer.buf[logBuffer.pos:logBuffer.pos+4], logBuffer.sizeBuf)
logBuffer.pos += size + 4
// Only now is the entry's window known: a rollover above seals the previous
// window first, and crediting this timestamp before that hands it to the
// sealed window and loses it from the new one - corrupting the received-ts
// eviction watermark in both directions.
if originalTsNs > logBuffer.curWindowMaxOriginalTsNs {
logBuffer.curWindowMaxOriginalTsNs = originalTsNs
}
logBuffer.offset++
return nil
}
func (logBuffer *LogBuffer) IsStopping() bool {
return logBuffer.isStopping.Load()
}
// ForceFlush immediately flushes the current buffer content and WAITS for completion
// This is useful for critical topics that need immediate persistence
// CRITICAL: This function is now SYNCHRONOUS - it blocks until the flush completes
func (logBuffer *LogBuffer) ForceFlush() {
if logBuffer.isStopping.Load() {
return // Don't flush if we're shutting down
}
logBuffer.Lock()
toFlush := logBuffer.copyToFlushWithCallback()
logBuffer.Unlock()
if toFlush != nil {
// The live buffer was already sealed and reset by copyToFlushWithCallback,
// so dropping toFlush on a timeout would lose it. Block until queued,
// bailing out only on shutdown.
if !logBuffer.queueFlush(toFlush) {
return
}
select {
case <-toFlush.done:
// Flush completed
case <-time.After(5 * time.Second):
// Queued but not yet flushed; loopFlush will still persist it
}
}
}
// ShutdownLogBuffer flushes the buffer and stops the log buffer
func (logBuffer *LogBuffer) ShutdownLogBuffer() {
isAlreadyStopped := logBuffer.isStopping.Swap(true)
if isAlreadyStopped {
return
}
// Wake any subscribers blocked in awaitNotificationOrTimeout so they can
// notice IsStopping() and exit promptly, even on an idle buffer where no
// flush notification would otherwise fire.
close(logBuffer.shutdownCh)
// Let go of the flush budget before sealing the last window, so a producer
// parked on it wakes up and the hand-off below cannot wait on a reservation.
logBuffer.flushBudget.close()
logBuffer.Lock()
toFlush := logBuffer.copyToFlush()
logBuffer.Unlock()
if toFlush != nil {
toFlush.budget = logBuffer.flushBudget.reserve(toFlush.seq, cap(toFlush.data))
logBuffer.flushChan <- toFlush
}
// nil is the shutdown sentinel: loopFlush drains everything queued before
// it and exits. The channel is never closed, so a sender racing shutdown
// can never panic on a closed channel.
logBuffer.flushChan <- nil
}
// IsAllFlushed returns true if all data in the buffer has been flushed, after calling ShutdownLogBuffer().
func (logBuffer *LogBuffer) IsAllFlushed() bool {
return logBuffer.isAllFlushed
}
// queueFlush hands a sealed window to loopFlush, reserving its bytes first so
// a producer waits for the queue to drain rather than adding another copy to
// it. Reports false when the buffer shut down before the hand-off.
func (logBuffer *LogBuffer) queueFlush(d *dataToFlush) bool {
// Charge the slab, not the window: mem.Allocate rounds up to a size class,
// so the bytes actually held are cap(data), and charging len would let the
// queue hold up to twice the ceiling.
d.budget = logBuffer.flushBudget.reserve(d.seq, cap(d.data))
// The window is already sealed, so dropping it here loses records the
// caller was told were accepted. Take any room in the queue first, and only
// fall back to the shutdown escape when there is none.
select {
case logBuffer.flushChan <- d:
return true
default:
}
select {
case logBuffer.flushChan <- d:
return true
case <-logBuffer.shutdownCh:
// shutting down; loopFlush may be gone, do not park forever
logBuffer.flushBudget.release(d.budget)
return false
}
}
func (logBuffer *LogBuffer) loopFlush() {
defer logBuffer.loopsDone.Done()
for d := range logBuffer.flushChan {
if d == nil {
break // shutdown sentinel
}
logBuffer.flushFn(logBuffer, d.startTime, d.stopTime, d.data, d.minOffset, d.maxOffset)
d.releaseMemory()
logBuffer.flushBudget.release(d.budget)
// local logbuffer is different from aggregate logbuffer here
if d.maxOffset >= 0 {
logBuffer.lastFlushedOffset.Store(d.maxOffset)
}
if !d.stopTime.IsZero() {
logBuffer.lastFlushTsNs.Store(d.stopTime.UnixNano())
}
// Wake readers that may be waiting to retry disk reads after the flush lands.
// LOAD-BEARING ORDER: the watermark store above must precede these
// notifications. A parked filer subscriber re-checks GetLastFlushTsNs on
// wake-up and goes back to sleep if it has not moved; notifying first
// opens a window where the wake-up looks spurious and the flush that
// caused it is only picked up by the retry timer. Not testable from
// outside (the window is nanoseconds on this goroutine) - keep the order.
if logBuffer.notifyFn != nil {
logBuffer.notifyFn()
}
logBuffer.notifySubscribers()
logBuffer.notifyFlushSubscribers()
// Signal completion if there's a callback channel
if d.done != nil {
close(d.done)
}
}
logBuffer.isAllFlushed = true
}
func (logBuffer *LogBuffer) loopInterval() {
defer logBuffer.loopsDone.Done()
// Wake on shutdown instead of sleeping through the interval: a goroutine
// parked in time.Sleep keeps the buffer and its ~40MB of slabs reachable
// for up to flushInterval after ShutdownLogBuffer.
ticker := time.NewTicker(logBuffer.flushInterval)
defer ticker.Stop()
for {
select {
case <-logBuffer.shutdownCh:
return
case <-ticker.C:
}
logBuffer.Lock()
toFlush := logBuffer.copyToFlush()
logBuffer.Unlock()
if toFlush != nil {
logBuffer.queueFlush(toFlush)
}
}
}
func (logBuffer *LogBuffer) copyToFlush() *dataToFlush {
return logBuffer.copyToFlushInternal(false)
}
func (logBuffer *LogBuffer) copyToFlushWithCallback() *dataToFlush {
return logBuffer.copyToFlushInternal(true)
}
func (logBuffer *LogBuffer) copyToFlushInternal(withCallback bool) *dataToFlush {
if logBuffer.pos > 0 {
var d *dataToFlush
if logBuffer.flushFn != nil {
d = &dataToFlush{
startTime: logBuffer.startTime,
stopTime: logBuffer.stopTime,
data: copiedBytes(logBuffer.buf[:logBuffer.pos]),
minOffset: logBuffer.minOffset,
maxOffset: logBuffer.maxOffset,
}
// Add callback channel for synchronous ForceFlush
if withCallback {
d.done = make(chan struct{})
}
// Stamped under the lock so the budget can admit windows in the
// order they were sealed. Every stamped window must reach reserve
// exactly once or the queue stalls behind the missing turn.
d.seq = logBuffer.flushSeq
logBuffer.flushSeq++
}
// CRITICAL: logBuffer.offset is the "next offset to assign", so last offset in buffer is offset-1
lastOffsetInBuffer := logBuffer.offset - 1
// Slot 0 falls out of the ring in SealBuffer below, so record how far
// eviction has reached before it goes - in both timestamp spaces.
if evicted := logBuffer.prevBuffers.buffers[0]; evicted.size > 0 && !evicted.stopTime.IsZero() {
if ts := evicted.stopTime.UnixNano(); ts > logBuffer.lastEvictedTsNs.Load() {
logBuffer.lastEvictedTsNs.Store(ts)
}
if ts := evicted.maxOriginalTsNs; ts > logBuffer.lastEvictedOriginalTsNs.Load() {
logBuffer.lastEvictedOriginalTsNs.Store(ts)
}
}
logBuffer.buf = logBuffer.prevBuffers.SealBuffer(logBuffer.startTime, logBuffer.stopTime, logBuffer.buf, logBuffer.pos, logBuffer.bufferStartOffset, lastOffsetInBuffer)
logBuffer.prevBuffers.buffers[len(logBuffer.prevBuffers.buffers)-1].maxOriginalTsNs = logBuffer.curWindowMaxOriginalTsNs
logBuffer.curWindowMaxOriginalTsNs = 0
// SealBuffer hands back the oldest window array to reuse. An entry larger
// than BufferSize grew one of these arrays to fit it, and buffers cycle
// forever, so without this a single oversized entry leaves every later
// window carrying — and snapshotting — its size. Growth is on demand, so
// the next oversized entry just reallocates.
if len(logBuffer.buf) > BufferSize {
logBuffer.buf = make([]byte, BufferSize)
}
// Hand a fully extended prefix snapshot to the sealed slot so sealed
// readers reuse it instead of re-copying the window; reset for the next
// window either way (holders keep their immutable prefix slices).
if len(logBuffer.curSnap) == logBuffer.pos {
logBuffer.prevBuffers.buffers[len(logBuffer.prevBuffers.buffers)-1].snapshot = logBuffer.curSnap[:logBuffer.pos:logBuffer.pos]
}
logBuffer.curSnap = nil
// Use zero time (time.Time{}) not epoch time (time.Unix(0,0))
// Epoch time (1970) breaks time-based reads after flush
logBuffer.startTime = time.Time{}
logBuffer.stopTime = time.Time{}
logBuffer.pos = 0
logBuffer.idx = logBuffer.idx[:0]
// DON'T increment offset - it's already pointing to the next offset!
// logBuffer.offset++ // REMOVED - this was causing offset gaps!
logBuffer.bufferStartOffset = logBuffer.offset // Next buffer starts at current offset (which is already the next one)
// Reset offset tracking
logBuffer.hasOffsets = false
logBuffer.minOffset = 0
logBuffer.maxOffset = 0
// Invalidate disk cache chunks after flush
// The cache may contain stale data from before this flush
// Invalidating ensures consumers will re-read fresh data from disk after flush
logBuffer.invalidateAllDiskCacheChunks()
return d
}
return nil
}
// invalidateAllDiskCacheChunks clears all cached disk chunks
// This should be called after a buffer flush to ensure consumers read fresh data from disk
func (logBuffer *LogBuffer) invalidateAllDiskCacheChunks() {
logBuffer.diskChunkCache.mu.Lock()
defer logBuffer.diskChunkCache.mu.Unlock()
if len(logBuffer.diskChunkCache.chunks) > 0 {
logBuffer.diskChunkCache.chunks = make(map[int64]*CachedDiskChunk)
}
}
// GetEarliestTime returns the oldest timestamp still resident in the buffer.
// It must consider the sealed prev buffers in addition to the active buffer,
// because ReadFromBuffer's tsMemory (and therefore ResumeFromDiskError) is
// computed from the min across both. Returning only the active startTime
// would cause gap-detection callers to skip past data still living in prev
// buffers, and can also silently equal the consumer's lastReadTime.
func (logBuffer *LogBuffer) GetEarliestTime() time.Time {
logBuffer.RLock()
defer logBuffer.RUnlock()
earliest := logBuffer.startTime
for _, prevBuf := range logBuffer.prevBuffers.buffers {
if prevBuf.startTime.IsZero() {
continue
}
if earliest.IsZero() || prevBuf.startTime.Before(earliest) {
earliest = prevBuf.startTime
}
}
return earliest
}
func (logBuffer *LogBuffer) HasData() bool {
logBuffer.RLock()
defer logBuffer.RUnlock()
if logBuffer.pos > 0 {
return true
}
for _, buf := range logBuffer.prevBuffers.buffers {
if buf.size > 0 {
return true
}
}
return false
}
func (logBuffer *LogBuffer) GetEarliestPosition() MessagePosition {
return MessagePosition{
Time: logBuffer.startTime,
Offset: logBuffer.offset,
}
}
// FlushedThroughTsNs reports the timestamp through which this buffer's data
// is durably on disk: "now" when nothing is pending a flush (a later append
// is bumped past the head), otherwise the last flushed window's stop time.
// Covers only entries that reached the buffer - callers stamping timestamps
// before the append must also bound by their in-flight floor (see
// Filer.LocalFlushedThroughTsNs).
func (logBuffer *LogBuffer) FlushedThroughTsNs(nowNs int64) int64 {
flushed := logBuffer.lastFlushTsNs.Load()
if logBuffer.LastTsNs.Load() <= flushed {
return nowNs
}
return flushed
}
// GetLastFlushTsNs returns the latest flushed timestamp in Unix nanoseconds.
// Returns 0 if nothing has been flushed yet.
func (logBuffer *LogBuffer) GetLastFlushTsNs() int64 {
return logBuffer.lastFlushTsNs.Load()
}
// GetLastEvictedOriginalTsNs is GetLastEvictedTsNs in pre-bump timestamps -
// the space disk cursors live in.
func (logBuffer *LogBuffer) GetLastEvictedOriginalTsNs() int64 {
return logBuffer.lastEvictedOriginalTsNs.Load()
}
// GetLastEvictedTsNs returns the stopTime of the newest window dropped from the
// sealed ring, or 0 if nothing has been evicted. A reader positioned past it
// knows the retained buffers still hold every entry after its position, which is
// the only emptiness proof available to a buffer that never flushes.
func (logBuffer *LogBuffer) GetLastEvictedTsNs() int64 {
return logBuffer.lastEvictedTsNs.Load()
}
// MarkEvictedThrough treats entries at or below tsNs as evicted even though
// nothing was rotated out yet: a merge-fed buffer is born empty while its
// sources hold history it must not skip. Call before the first append.
func (logBuffer *LogBuffer) MarkEvictedThrough(tsNs int64) {
if tsNs > logBuffer.lastEvictedTsNs.Load() {
logBuffer.lastEvictedTsNs.Store(tsNs)
}
if tsNs > logBuffer.lastEvictedOriginalTsNs.Load() {
logBuffer.lastEvictedOriginalTsNs.Store(tsNs)
}
}
func (logBuffer *LogBuffer) SetLastFlushTsNs(ts int64) {
logBuffer.lastFlushTsNs.Store(ts)
}
func (d *dataToFlush) releaseMemory() {
// Guard nil: mem.Free(nil) would put a zero-cap slice into the smallest slot
// pool, so a later Allocate could hand back nil and panic. Also makes a double
// release harmless.
if d.data != nil {
mem.Free(d.data)
d.data = nil
}
}
// ReadFromBuffer returns the in-memory log data at lastReadPosition. isPooled
// reports whether the returned buffer is a private pooled copy the caller must
// return via ReleaseMemory; when false the buffer wraps a snapshot shared with
// other readers and must not be released (or written to).
func (logBuffer *LogBuffer) ReadFromBuffer(lastReadPosition MessagePosition) (bufferCopy *bytes.Buffer, batchIndex int64, isPooled bool, err error) {
logBuffer.RLock()
defer logBuffer.RUnlock()
isOffsetBased := lastReadPosition.IsOffsetBased
// For offset-based subscriptions, use offset comparisons, not time comparisons!
if isOffsetBased {
requestedOffset := lastReadPosition.Offset
// Check if the requested offset is in the current buffer range
if requestedOffset >= logBuffer.bufferStartOffset && requestedOffset <= logBuffer.offset {
// If current buffer is empty (pos=0), check if data is on disk or not yet written
if logBuffer.pos == 0 {
// If buffer is empty but offset range covers the request,
// it means data was in memory and has been flushed/moved out.
// The bufferStartOffset advancing to cover this offset proves data existed.
//
// Three cases:
// 1. requestedOffset < logBuffer.offset: Data was here, now flushed
// 2. requestedOffset == logBuffer.offset && bufferStartOffset > 0: Buffer advanced, data flushed
// 3. requestedOffset == logBuffer.offset && bufferStartOffset == 0: Initial state - try disk first!
//
// Cases 1 & 2: try disk read
// Case 3: try disk read (historical data might exist)
if requestedOffset < logBuffer.offset {
// Data was in the buffer range but buffer is now empty = flushed to disk
return nil, -2, false, ResumeFromDiskError
}
// requestedOffset == logBuffer.offset: Current position
// CRITICAL: For subscribers starting from offset 0, try disk read first
// (historical data might exist from previous runs)
if requestedOffset == 0 && logBuffer.bufferStartOffset == 0 && logBuffer.offset == 0 {
// Initial state: try disk read before waiting for new data
return nil, -2, false, ResumeFromDiskError
}
// Otherwise, wait for new data to arrive
return nil, logBuffer.offset, false, nil
}
return logBuffer.currentSnapshotView(0, logBuffer.pos), logBuffer.offset, false, nil
}
// Check previous buffers for the requested offset
for _, buf := range logBuffer.prevBuffers.buffers {
if requestedOffset >= buf.startOffset && requestedOffset <= buf.offset {
// If prevBuffer is empty, it means the data was flushed to disk
// (prevBuffers are created when buffer is flushed)
if buf.size == 0 {
// Empty prevBuffer covering this offset means data was flushed
return nil, -2, false, ResumeFromDiskError
}
return sharedBufferView(buf, 0), buf.offset, false, nil
}
}
// Offset not found in any buffer
if requestedOffset < logBuffer.bufferStartOffset {
// Data not in current buffers - must be on disk (flushed or never existed)
// Return ResumeFromDiskError to trigger disk read
return nil, -2, false, ResumeFromDiskError
}
if requestedOffset > logBuffer.offset {
// Future data, not available yet
return nil, logBuffer.offset, false, nil
}
// Offset not found - return nil
return nil, logBuffer.offset, false, nil
}
// TIMESTAMP-BASED READ (original logic)
// Read from disk and memory
// 1. read from disk, last time is = td
// 2. in memory, the earliest time = tm
// if tm <= td, case 2.1
// read from memory
// if tm is empty, case 2.2
// read from memory
// if td < tm, case 2.3
// read from disk again
var tsMemory time.Time
if !logBuffer.startTime.IsZero() {
tsMemory = logBuffer.startTime
}
for _, prevBuf := range logBuffer.prevBuffers.buffers {
if !prevBuf.startTime.IsZero() {
// If tsMemory is zero, assign directly; otherwise compare
if tsMemory.IsZero() || prevBuf.startTime.Before(tsMemory) {
tsMemory = prevBuf.startTime
}
}
}
if tsMemory.IsZero() { // case 2.2
// Buffer is empty - return ResumeFromDiskError so caller can read from disk
// This fixes issue #4977 where SubscribeMetadata stalls because
// MetaAggregator.MetaLogBuffer is empty in single-filer setups
return nil, -2, false, ResumeFromDiskError
} else if lastReadPosition.Time.Before(tsMemory) { // case 2.3
// For time-based reads, only check timestamp for disk reads
// Don't use offset comparisons as they're not meaningful for time-based subscriptions
// A gated cursor below the eviction watermark must go to disk: serving it
// from the earliest retained window would silently skip the evicted span.
if lastReadPosition.Offset == EvictionGatedOffset && lastReadPosition.Time.UnixNano() < logBuffer.lastEvictedTsNs.Load() {
return nil, -2, false, ResumeFromDiskError
}
// Special case: If requested time is zero (Unix epoch), treat as "start from beginning"
// This handles queries that want to read all data without knowing the exact start time
if lastReadPosition.Time.IsZero() || lastReadPosition.Time.Unix() == 0 {
// Start from the beginning of memory
// Fall through to case 2.1 to read from earliest buffer
} else if lastReadPosition.Offset <= 0 && lastReadPosition.Time.Before(tsMemory) {
// Treat first read with sentinel/zero offset as inclusive of earliest in-memory data
} else {
// Data not in memory buffers - read from disk
return nil, -2, false, ResumeFromDiskError
}
}
// the following is case 2.1
if lastReadPosition.Time.Equal(logBuffer.stopTime) && !logBuffer.stopTime.IsZero() {
// For first-read sentinel/zero offset, allow inclusive read at the boundary
if lastReadPosition.Offset > 0 {
return nil, logBuffer.offset, false, nil
}
}
if lastReadPosition.Time.After(logBuffer.stopTime) && !logBuffer.stopTime.IsZero() {
return nil, logBuffer.offset, false, nil
}
// Also check prevBuffers when current buffer is empty (startTime is zero)
if lastReadPosition.Time.Before(logBuffer.startTime) || logBuffer.startTime.IsZero() {
for _, buf := range logBuffer.prevBuffers.buffers {
if buf.startTime.After(lastReadPosition.Time) {
return sharedBufferView(buf, 0), buf.offset, false, nil
}
if !buf.startTime.After(lastReadPosition.Time) && buf.stopTime.After(lastReadPosition.Time) {
searchTime := lastReadPosition.Time
if lastReadPosition.Offset <= 0 {
searchTime = searchTime.Add(-time.Nanosecond)
}
pos, err := buf.locateByTs(searchTime)
if err != nil {
// Buffer corruption detected - return error wrapped with ErrBufferCorrupted
glog.Errorf("ReadFromBuffer: buffer corruption in prevBuffer: %v", err)
return nil, -1, false, fmt.Errorf("%w: %v", ErrBufferCorrupted, err)
}
if pos < buf.size {
return sharedBufferView(buf, pos), buf.offset, false, nil
}
}
}
// If current buffer is not empty, return it
if logBuffer.pos > 0 {
return logBuffer.currentSnapshotView(0, logBuffer.pos), logBuffer.offset, false, nil
}
// Buffer is empty and no data in prevBuffers - wait for new data
return nil, logBuffer.offset, false, nil
}
lastTs := lastReadPosition.Time.UnixNano()
// Inclusive boundary for first-read sentinel/zero offset
searchTs := lastTs
if lastReadPosition.Offset <= 0 {
if searchTs > math.MinInt64+1 { // prevent underflow
searchTs = searchTs - 1
}
}
l, h := 0, len(logBuffer.idx)-1
/*
for i, pos := range m.idx {
logEntry, ts := readTs(m.buf, pos)
event := &filer_pb.SubscribeMetadataResponse{}
proto.Unmarshal(logEntry.Data, event)
entry := event.EventNotification.OldEntry
if entry == nil {
entry = event.EventNotification.NewEntry
}
}
*/
for l <= h {
mid := (l + h) / 2
pos := logBuffer.idx[mid]
_, t, err := readTs(logBuffer.buf, pos)
if err != nil {
// Buffer corruption detected in binary search
glog.Errorf("ReadFromBuffer: buffer corruption at idx[%d] pos %d: %v", mid, pos, err)
return nil, -1, false, fmt.Errorf("%w: %v", ErrBufferCorrupted, err)
}
if t <= searchTs {
l = mid + 1
} else if searchTs < t {
var prevT int64
if mid > 0 {
_, prevT, err = readTs(logBuffer.buf, logBuffer.idx[mid-1])
if err != nil {
// Buffer corruption detected in binary search (previous entry)
glog.Errorf("ReadFromBuffer: buffer corruption at idx[%d] pos %d: %v", mid-1, logBuffer.idx[mid-1], err)
return nil, -1, false, fmt.Errorf("%w: %v", ErrBufferCorrupted, err)
}
}
if prevT <= searchTs {
return logBuffer.currentSnapshotView(pos, logBuffer.pos), logBuffer.offset, false, nil
}
h = mid
}
}
// Binary search didn't find the timestamp - data may have been flushed to disk already
// Returning -2 signals to caller that data is not available in memory
return nil, -2, false, nil
}
// ReleaseMemory returns a pooled buffer for reuse. Only call it for buffers
// ReadFromBuffer reported as pooled: recycling a shared sealed-window view
// would let the next copiedBytes overwrite bytes other readers still hold.
func (logBuffer *LogBuffer) ReleaseMemory(b *bytes.Buffer) {
bufferPool.Put(b)
}
// GetName returns the log buffer name for metadata tracking
func (logBuffer *LogBuffer) GetName() string {
logBuffer.RLock()
defer logBuffer.RUnlock()
return logBuffer.name
}
// GetOffset returns the current offset for metadata tracking
func (logBuffer *LogBuffer) GetOffset() int64 {
logBuffer.RLock()
defer logBuffer.RUnlock()
return logBuffer.offset
}
var bufferPool = sync.Pool{
New: func() interface{} {
return new(bytes.Buffer)
},
}
// copiedBytes returns a private copy of buf backed by the shared, size-classed
// slab pool. The caller owns it until mem.Free (see dataToFlush.releaseMemory),
// so the live buffer never outlives the flush. Routing through mem reuses slabs
// across the process instead of reallocating a window-sized array per flush.
func copiedBytes(buf []byte) []byte {
copied := mem.Allocate(len(buf))
copy(copied, buf)
return copied
}
// sharedBufferView wraps the sealed window's shared snapshot from pos onward.
// The returned buffer aliases memory shared with other readers: it must be
// treated as read-only and never passed to ReleaseMemory (the full-slice cap
// makes an accidental append reallocate instead of scribbling on the snapshot).
func sharedBufferView(mb *MemBuffer, pos int) *bytes.Buffer {
snap := mb.sharedSnapshot()
return bytes.NewBuffer(snap[pos:len(snap):len(snap)])
}
// currentSnapshotView returns buf[from:to] of the current window as a view of
// the shared prefix snapshot, extending the snapshot to cover [0:to) first.
// buf[:pos] is append-only until the window seals, so extension only ever
// appends stable bytes; earlier holders' slices are unaffected. Callers must
// hold the read lock (keeps buf and pos stable during extension).
func (logBuffer *LogBuffer) currentSnapshotView(from, to int) *bytes.Buffer {
logBuffer.curSnapMu.Lock()
if cap(logBuffer.curSnap) < to {
// First use in this window (or the rare window whose array outgrew the
// previous one): size to the window array so extensions never reallocate.
grown := make([]byte, len(logBuffer.curSnap), len(logBuffer.buf))
copy(grown, logBuffer.curSnap)
logBuffer.curSnap = grown
}
if len(logBuffer.curSnap) < to {
logBuffer.curSnap = append(logBuffer.curSnap, logBuffer.buf[len(logBuffer.curSnap):to]...)
}
snap := logBuffer.curSnap[:to]
logBuffer.curSnapMu.Unlock()
return bytes.NewBuffer(snap[from:to:to])
}
func readTs(buf []byte, pos int) (size int, ts int64, err error) {
// Bounds check for size field (overflow-safe)
if pos < 0 || pos > len(buf)-4 {
return 0, 0, fmt.Errorf("corrupted log buffer: cannot read size at pos %d, buffer length %d", pos, len(buf))
}
size = int(util.BytesToUint32(buf[pos : pos+4]))
// Bounds check for entry data (overflow-safe, protects against negative size)
if size < 0 || size > len(buf)-pos-4 {
return 0, 0, fmt.Errorf("corrupted log buffer: entry size %d at pos %d exceeds buffer length %d", size, pos, len(buf))
}
entryData := buf[pos+4 : pos+4+size]
// Read only LogEntry.ts_ns rather than unmarshaling the whole entry. This
// runs on every binary-search probe in ReadFromBuffer; a full proto.Unmarshal
// there allocates fresh slices for the data/key byte fields on each call, which
// dominated allocation churn under metadata-subscription fan-out.
ts, err = readTsNs(entryData)
if err != nil {
return 0, 0, fmt.Errorf("corrupted log buffer at pos %d, size %d: %w", pos, size, err)
}
return size, ts, nil
}
// readTsNs scans a marshaled LogEntry and returns its ts_ns (field 1, varint)
// without decoding the data/key byte fields. Fields serialize in number order,
// so ts_ns is normally the first tag and this returns after one varint; the full
// scan keeps it correct for any field order. A missing field 1 means the proto3
// default, ts_ns == 0.
func readTsNs(entryData []byte) (tsNs int64, err error) {
for i := 0; i < len(entryData); {
tag, n := binary.Uvarint(entryData[i:])
if n <= 0 {
return 0, fmt.Errorf("bad field tag at %d", i)
}
i += n
fieldNum := tag >> 3
switch tag & 0x7 { // wire type
case 0: // varint
v, m := binary.Uvarint(entryData[i:])
if m <= 0 {
return 0, fmt.Errorf("bad varint for field %d", fieldNum)
}
i += m
if fieldNum == 1 { // ts_ns
return int64(v), nil
}
case 1: // 64-bit
i += 8
case 2: // length-delimited: read length, skip payload without copying
l, m := binary.Uvarint(entryData[i:])
if m <= 0 {
return 0, fmt.Errorf("bad length for field %d", fieldNum)
}
i += m
if l > uint64(len(entryData)-i) {
return 0, fmt.Errorf("field %d length %d overruns buffer", fieldNum, l)
}
i += int(l)
case 5: // 32-bit
i += 4
default:
return 0, fmt.Errorf("unknown wire type for field %d", fieldNum)
}
if i > len(entryData) {
return 0, fmt.Errorf("field %d overruns buffer", fieldNum)
}
}
return 0, nil
}
// unmarshalLogEntryAliased decodes a marshaled LogEntry into out, pointing the
// data and key fields at sub-slices of entryData instead of copying them. A full
// proto.Unmarshal allocates fresh slices for those byte fields on every entry
// (protobuf consumeBytesNoZero), which dominated allocation churn when many
// metadata subscribers each re-read the in-memory log window.
//
// The aliased data/key are only valid while entryData is, i.e. for the duration
// of the eachLogEntryFn callback. Callers must copy anything they retain past the
// callback; all current subscribers do (they re-decode data into their own event
// or hand it to a synchronous grpc Send).
func unmarshalLogEntryAliased(entryData []byte, out *filer_pb.LogEntry) error {
out.TsNs = 0
out.PartitionKeyHash = 0
out.Data = nil
out.Key = nil
out.Offset = 0
for i := 0; i < len(entryData); {
tag, n := binary.Uvarint(entryData[i:])
if n <= 0 {
return fmt.Errorf("bad field tag at %d", i)
}
i += n
fieldNum := tag >> 3
switch tag & 0x7 { // wire type
case 0: // varint: ts_ns / partition_key_hash / offset
v, m := binary.Uvarint(entryData[i:])
if m <= 0 {
return fmt.Errorf("bad varint for field %d", fieldNum)
}
i += m
switch fieldNum {
case 1:
out.TsNs = int64(v)
case 2:
out.PartitionKeyHash = int32(v)
case 5:
out.Offset = int64(v)
}
case 2: // length-delimited: data / key, aliased not copied
l, m := binary.Uvarint(entryData[i:])
if m <= 0 {
return fmt.Errorf("bad length for field %d", fieldNum)
}
i += m
if l > uint64(len(entryData)-i) {
return fmt.Errorf("field %d length %d overruns buffer", fieldNum, l)
}
switch fieldNum {
case 3:
out.Data = entryData[i : i+int(l)]
case 4:
out.Key = entryData[i : i+int(l)]
}
i += int(l)
case 1: // 64-bit
i += 8
case 5: // 32-bit
i += 4
default:
return fmt.Errorf("unknown wire type for field %d", fieldNum)
}
if i > len(entryData) {
return fmt.Errorf("field %d overruns buffer", fieldNum)
}
}
return nil
}