Replace RefreshSafeTail with epoch table based safe tail tracking [dev]

[badrishc]

No description provided.

[Copilot]

Pull request overview

This PR replaces the previous “background refresh / refresh-frequency” SafeTail mechanism in TsavoriteLog with a lazy, epoch-table-based SafeTail computation (via LightEpoch user-word slots), and removes the associated tuning knobs from settings/options across the repo.

Changes:

• Introduces LightEpoch “user-word” slots and uses them in TsavoriteLog to track per-thread in-flight enqueue lower bounds for SafeTail computation.
• Removes SafeTailRefreshFrequencyMs and related Garnet config/options, updating AOF/pubsub construction accordingly.
• Updates log-scan iterators and tests to use RefreshSafeTailAddress() and cached SafeTail reads instead of relying on periodic refresh.

Reviewed changes

Copilot reviewed 14 out of 14 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
libs/storage/Tsavorite/cs/test/LogTests.cs	Updates tests to wait using RefreshSafeTailAddress() and removes deprecated settings usage.
libs/storage/Tsavorite/cs/test/LogScanTests.cs	Removes SafeTailRefreshFrequencyMs from test log settings and refreshes safe tail via new API.
libs/storage/Tsavorite/cs/src/core/TsavoriteLog/TsavoriteLogSettings.cs	Removes SafeTailRefreshFrequencyMs setting.
libs/storage/Tsavorite/cs/src/core/TsavoriteLog/TsavoriteLogScanSingleIterator.cs	Adds per-iterator backoff around safe-tail refresh scans for single-iterator waits.
libs/storage/Tsavorite/cs/src/core/TsavoriteLog/TsavoriteLogScanIterator.cs	Uses cached SafeTailAddress on hot paths; refreshes safe tail in uncommitted wait path.
libs/storage/Tsavorite/cs/src/core/TsavoriteLog/TsavoriteLog.cs	Implements epoch-user-word-based SafeTail caching/refresh and replaces old callback/worker refresh logic.
libs/storage/Tsavorite/cs/src/core/Epochs/LightEpoch.cs	Adds user-word slot allocation, per-thread access, and min-fold scanning APIs.
libs/server/Servers/GarnetServerOptions.cs	Removes AOF replication refresh-frequency option and stops passing it into TsavoriteLogSettings.
libs/server/PubSub/SubscribeBroker.cs	Removes subscriber refresh-frequency constructor parameter and setting usage.
libs/host/defaults.conf	Removes removed refresh-frequency configuration keys.
libs/host/GarnetServer.cs	Updates SubscribeBroker construction to new signature (no refresh-frequency).
libs/host/Configuration/Options.cs	Removes CLI options for AOF/subscriber refresh-frequency and corresponding mapping.
libs/cluster/Server/Replication/PrimaryOps/AofTaskStore.cs	Renames and simplifies safe-tail page-shift callback usage for fast truncation.
benchmark/BDN.benchmark/Embedded/EmbeddedRespServer.cs	Updates SubscribeBroker construction to new signature.

Comments suppressed due to low confidence (1)

libs/storage/Tsavorite/cs/src/core/TsavoriteLog/TsavoriteLogScanSingleIterator.cs:92

• The backoff gate can cause the iterator to wait even though data is already available: if scanBackoffUntil is in the future (after previous failed scans) and a producer enqueues a single record then stops, the onEnqueue signal wakes the iterator but it may skip RefreshSafeTailAddress() due to backoff and then await onEnqueue again, potentially blocking forever. Reset scanBackoffUntil (and/or bypass backoff) on enqueue signals and when resetting scanBackoffTicks to 0 so a wake-up always leads to at least one safe-tail refresh attempt.

                // Fast check: cached SafeTailAddress may already be ahead (e.g., another iterator
                // or the multi-iterator pre-refresh in NotifyParkedWaiters advanced it).
                if (this.NextAddress < this.tsavoriteLog.SafeTailAddress)
                {
                    scanBackoffTicks = 0; // data available — reset backoff
                    return true;
                }

                // Scan the epoch table only if the backoff interval has elapsed. This limits scan
                // frequency when the consumer is caught up and the producer is slow, capping at one
                // scan per 10ms in the worst case (matching the old baseline refresh frequency).
                if (Stopwatch.GetTimestamp() >= scanBackoffUntil)
                {
                    if (this.NextAddress < this.tsavoriteLog.RefreshSafeTailAddress())
                    {
                        scanBackoffTicks = 0; // scan advanced past NextAddress — reset backoff
                        return true;
                    }

                    // Scan didn't help — double the backoff interval (starting from ~100μs).
                    scanBackoffTicks = scanBackoffTicks == 0
                        ? Stopwatch.Frequency / 10_000  // initial: ~100μs
                        : Math.Min(scanBackoffTicks * 2, MaxBackoffTicks);
                    scanBackoffUntil = Stopwatch.GetTimestamp() + scanBackoffTicks;
                }

                // Ignore refresh-uncommitted exceptions, except when the token is signaled
                await onEnqueue.WaitAsync().ConfigureAwait(false);
            }

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.