| | | 1 | | using Microsoft.Extensions.Hosting; |
| | | 2 | | using Microsoft.Extensions.Logging; |
| | | 3 | | using Microsoft.Extensions.Options; |
| | | 4 | | using System.Diagnostics; |
| | | 5 | | |
| | | 6 | | namespace AsyncResponse; |
| | | 7 | | |
| | | 8 | | /// <summary>Options for the async-response recovery watchdog.</summary> |
| | | 9 | | public sealed class AsyncResponseWatchdogOptions |
| | | 10 | | { |
| | | 11 | | /// <summary> |
| | | 12 | | /// Whether the watchdog runs. Default: <c>true</c>. Set to <c>false</c> to disable it — for |
| | | 13 | | /// example in all but one host when several hosts share one durable recovery store, so the |
| | | 14 | | /// scan and its warnings are not duplicated. |
| | | 15 | | /// </summary> |
| | | 16 | | public bool Enabled { get; set; } = true; |
| | | 17 | | |
| | | 18 | | /// <summary>How often the watchdog scans the persisted recovery state. Default: 6 hours.</summary> |
| | | 19 | | public TimeSpan Interval { get; set; } = TimeSpan.FromHours(6); |
| | | 20 | | |
| | | 21 | | /// <summary> |
| | | 22 | | /// Age past which a recovery entry with no live subscriber is reported as stale. |
| | | 23 | | /// Default: 24 hours. |
| | | 24 | | /// </summary> |
| | | 25 | | public TimeSpan StaleAfter { get; set; } = TimeSpan.FromHours(24); |
| | | 26 | | |
| | | 27 | | /// <summary>Delay before the first scan, so startup is never blocked. Default: 5 minutes.</summary> |
| | | 28 | | public TimeSpan StartupDelay { get; set; } = TimeSpan.FromMinutes(5); |
| | | 29 | | |
| | | 30 | | /// <summary> |
| | | 31 | | /// Upper bound on the random extra delay added to the startup delay and to every interval |
| | | 32 | | /// wait. Default: 10% of <see cref="Interval"/>. |
| | | 33 | | /// <para> |
| | | 34 | | /// Replicas deployed together start together, so a fixed interval keeps them scanning in |
| | | 35 | | /// lockstep forever: every host walks the same recovery store and fires its own liveness probe |
| | | 36 | | /// per correlation id at the same instant, turning a routine scan into a synchronized burst |
| | | 37 | | /// against the channel. An independent offset per replica spreads that out and costs nothing — |
| | | 38 | | /// the scan is a periodic report, so when it runs within the interval does not matter. |
| | | 39 | | /// </para> |
| | | 40 | | /// <para> |
| | | 41 | | /// Set to <see cref="TimeSpan.Zero"/> for an exactly-periodic scan (single-host deployments, |
| | | 42 | | /// or tests that assert on scan timing). |
| | | 43 | | /// </para> |
| | | 44 | | /// </summary> |
| | | 45 | | public TimeSpan? IntervalJitter { get; set; } |
| | | 46 | | |
| | | 47 | | /// <summary>The resolved jitter bound: <see cref="IntervalJitter"/> or 10% of the interval.</summary> |
| | | 48 | | internal TimeSpan ResolvedJitter => IntervalJitter ?? TimeSpan.FromTicks(Interval.Ticks / 10); |
| | | 49 | | |
| | | 50 | | /// <summary> |
| | | 51 | | /// Upper bound on the recovery entries one scan buffers (the scan dedupes in memory before |
| | | 52 | | /// probing liveness). When the store holds more, the scan stops enumerating at the cap, |
| | | 53 | | /// reports the buffered subset (<see cref="AsyncResponseWatchdogReport.Truncated"/> is set, |
| | | 54 | | /// the health check degrades), and logs a warning — bounding scan memory on very large |
| | | 55 | | /// stores at the cost of an incomplete staleness report. The count is a MEMORY bound, not a |
| | | 56 | | /// flow count: grouped entries occupy one slot per unique correlation id, correlation-less |
| | | 57 | | /// entries one slot per row. Default: 100 000. |
| | | 58 | | /// </summary> |
| | | 59 | | public int MaxScanEntries { get; set; } = 100_000; |
| | | 60 | | |
| | | 61 | | /// <summary> |
| | | 62 | | /// Upper bound on liveness probes one scan runs concurrently. Each probe is its own round |
| | | 63 | | /// trip to the channel (a store query or broker request) and a scan issues one per buffered |
| | | 64 | | /// entry, so probing strictly sequentially would serialize up to <see cref="MaxScanEntries"/> |
| | | 65 | | /// round trips per scan. Must be at least 1 (strictly sequential). Default: 8. |
| | | 66 | | /// </summary> |
| | | 67 | | public int ProbeConcurrency { get; set; } = 8; |
| | | 68 | | |
| | | 69 | | /// <summary>Validates the scan-loop knobs so a bad value fails at startup, not mid-scan.</summary> |
| | | 70 | | internal void Validate() |
| | | 71 | | { |
| | | 72 | | // Interval and StartupDelay arm Task.Delay in the scan loop; an out-of-range value there |
| | | 73 | | // would throw outside the per-scan try, fault the background service mid-run, and (with |
| | | 74 | | // the default BackgroundServiceExceptionBehavior.StopHost) take the host down instead of |
| | | 75 | | // failing fast where the misconfiguration is visible. |
| | | 76 | | AsyncResponseChannelOptions.EnsureTimerBacked(Interval, nameof(AsyncResponseWatchdogOptions), nameof(Interval)); |
| | | 77 | | AsyncResponseChannelOptions.EnsureTimerBackedAllowZero(StartupDelay, nameof(AsyncResponseWatchdogOptions), nameo |
| | | 78 | | AsyncResponseChannelOptions.EnsureTimerBackedAllowZero(ResolvedJitter, nameof(AsyncResponseWatchdogOptions), nam |
| | | 79 | | |
| | | 80 | | // Validating the parts separately is not enough, because NextWait arms their SUM: two |
| | | 81 | | // individually legal values can still hand Task.Delay an out-of-range delay and fault the |
| | | 82 | | // background service on its very first wait — the exact delayed, host-stopping failure |
| | | 83 | | // these checks exist to prevent. |
| | | 84 | | AsyncResponseChannelOptions.EnsureTimerBackedAllowZero( |
| | | 85 | | StartupDelay + ResolvedJitter, nameof(AsyncResponseWatchdogOptions), $"{nameof(StartupDelay)} + {nameof(Inte |
| | | 86 | | AsyncResponseChannelOptions.EnsureTimerBackedAllowZero( |
| | | 87 | | Interval + ResolvedJitter, nameof(AsyncResponseWatchdogOptions), $"{nameof(Interval)} + {nameof(IntervalJitt |
| | | 88 | | |
| | | 89 | | // Jitter is a de-synchronization offset, not a second interval. Allowing it to exceed the |
| | | 90 | | // interval lets a perfectly healthy next scan be scheduled beyond the freshness budget the |
| | | 91 | | // recovery health check derives from Interval, so readiness reports the watchdog dead while |
| | | 92 | | // it is doing exactly what it was configured to do. |
| | | 93 | | if (ResolvedJitter > Interval) |
| | | 94 | | { |
| | | 95 | | throw new InvalidOperationException( |
| | | 96 | | $"{nameof(AsyncResponseWatchdogOptions)}.{nameof(IntervalJitter)} ({ResolvedJitter}) cannot exceed " + |
| | | 97 | | $"{nameof(Interval)} ({Interval}): the health check derives its freshness budget from the interval, so a |
| | | 98 | | "larger offset would report a healthy watchdog as stale."); |
| | | 99 | | } |
| | | 100 | | |
| | | 101 | | // The probe fan-out degree feeds Parallel.ForEachAsync mid-scan, which rejects values |
| | | 102 | | // below 1 with the same delayed, host-stopping failure mode. |
| | | 103 | | if (ProbeConcurrency < 1) |
| | | 104 | | throw new InvalidOperationException( |
| | | 105 | | $"{nameof(AsyncResponseWatchdogOptions)}.{nameof(ProbeConcurrency)} must be at least 1."); |
| | | 106 | | |
| | | 107 | | // The buffer cap gates growth with ">= MaxScanEntries", so a non-positive value truncates |
| | | 108 | | // on the FIRST entry: every scan then classifies an empty set and publishes a report that |
| | | 109 | | // is Truncated with zeroed counters. The health check degrades permanently and no stale |
| | | 110 | | // registration is ever reported — the stuck-flow alarm is silently off while the logs |
| | | 111 | | // still show a scan completing each interval. Nothing later in the scan can catch this, |
| | | 112 | | // so it has to fail here. ("Unlimited" is not a supported value: the cap is a memory |
| | | 113 | | // bound, and int.MaxValue is the way to ask for effectively no limit.) |
| | | 114 | | if (MaxScanEntries < 1) |
| | | 115 | | throw new InvalidOperationException( |
| | | 116 | | $"{nameof(AsyncResponseWatchdogOptions)}.{nameof(MaxScanEntries)} must be at least 1; it bounds scan mem |
| | | 117 | | $"so use {int.MaxValue} for effectively no limit rather than zero."); |
| | | 118 | | |
| | | 119 | | // Staleness is judged as "utcNow - registeredAtUtc >= StaleAfter", so a non-positive |
| | | 120 | | // threshold makes EVERY live registration stale and logs a Warning per entry on every |
| | | 121 | | // scan — an alarm that fires constantly is the same as no alarm at all. |
| | | 122 | | if (StaleAfter <= TimeSpan.Zero) |
| | | 123 | | throw new InvalidOperationException( |
| | | 124 | | $"{nameof(AsyncResponseWatchdogOptions)}.{nameof(StaleAfter)} must be positive; a non-positive threshold |
| | | 125 | | "every live registration as stale."); |
| | | 126 | | } |
| | | 127 | | } |
| | | 128 | | |
| | | 129 | | /// <summary> |
| | | 130 | | /// Snapshot of one persisted recovery entry as observed by the watchdog. |
| | | 131 | | /// </summary> |
| | | 132 | | /// <param name="CorrelationId">The correlation id the entry belongs to.</param> |
| | | 133 | | /// <param name="RegisteredAtUtc">When the waiter registered, or <c>null</c> if unknown.</param> |
| | | 134 | | /// <param name="ActiveSubscribers"> |
| | | 135 | | /// Live subscribers awaiting this correlation id's channel: <c>0</c> = no live waiter, a positive |
| | | 136 | | /// value = at least one, a negative value = liveness could not be probed (no |
| | | 137 | | /// <see cref="IActiveSubscriberProbe"/>). |
| | | 138 | | /// </param> |
| | | 139 | | /// <param name="PayloadTypeFullName">The payload type the waiter subscribed for.</param> |
| | | 140 | | public sealed record RecoveryStateObservation( |
| | | 141 | | string? CorrelationId, |
| | | 142 | | DateTime? RegisteredAtUtc, |
| | | 143 | | long ActiveSubscribers, |
| | | 144 | | string? PayloadTypeFullName); |
| | | 145 | | |
| | | 146 | | /// <summary> |
| | | 147 | | /// Outcome of one watchdog scan attempt, as published for consumers |
| | | 148 | | /// (e.g. <see cref="AsyncResponseRecoveryHealthCheck"/>). |
| | | 149 | | /// </summary> |
| | | 150 | | /// <param name="ScanCompletedUtc">When the scan attempt finished.</param> |
| | | 151 | | /// <param name="ScanInterval">The configured scan interval, so consumers can judge snapshot freshness.</param> |
| | | 152 | | /// <param name="Report">The evaluation result; <c>null</c> when the scan failed.</param> |
| | | 153 | | /// <param name="Error">The scan failure message; <c>null</c> when the scan succeeded.</param> |
| | | 154 | | public sealed record AsyncResponseWatchdogSnapshot( |
| | | 155 | | DateTime ScanCompletedUtc, |
| | | 156 | | TimeSpan ScanInterval, |
| | | 157 | | AsyncResponseWatchdogReport? Report, |
| | | 158 | | string? Error); |
| | | 159 | | |
| | | 160 | | /// <summary> |
| | | 161 | | /// Holds the latest watchdog scan result. The watchdog is the single writer; readers (e.g. the |
| | | 162 | | /// readiness health check) get a cheap, consistent snapshot without touching the recovery store. |
| | | 163 | | /// </summary> |
| | | 164 | | public sealed class AsyncResponseWatchdogState |
| | | 165 | | { |
| | | 166 | | private volatile AsyncResponseWatchdogSnapshot? _latest; |
| | | 167 | | private volatile WatchdogActivation? _activation; |
| | | 168 | | |
| | | 169 | | /// <summary>The most recent scan outcome, or <c>null</c> when no scan has completed yet.</summary> |
| | 11873 | 170 | | public AsyncResponseWatchdogSnapshot? Latest => _latest; |
| | | 171 | | |
| | | 172 | | /// <summary> |
| | | 173 | | /// Whether this host's watchdog scans: <c>true</c> once its scan loop is armed, <c>false</c> |
| | | 174 | | /// when it declined to run (disabled via options, or the channel registers no scanner), and |
| | | 175 | | /// <c>null</c> while unknown (the watchdog has not started, or none is registered). |
| | | 176 | | /// </summary> |
| | 6 | 177 | | public bool? Scanning => _activation?.Scanning; |
| | | 178 | | |
| | | 179 | | /// <summary>Why this host's watchdog does not scan, when <see cref="Scanning"/> is <c>false</c>.</summary> |
| | 6 | 180 | | public string? IdleReason => _activation is { Scanning: false } activation ? activation.IdleReason : null; |
| | | 181 | | |
| | 41 | 182 | | internal WatchdogActivation? Activation => _activation; |
| | | 183 | | |
| | | 184 | | /// <summary>Publishes the latest watchdog snapshot for health checks and metrics.</summary> |
| | 1637 | 185 | | public void Publish(AsyncResponseWatchdogSnapshot snapshot) => _latest = snapshot; |
| | | 186 | | |
| | | 187 | | /// <summary> |
| | | 188 | | /// Marks this host's watchdog as armed, so the health check can hold it to a first-scan |
| | | 189 | | /// deadline instead of attesting "no scan yet" for a loop that died before ever publishing. |
| | | 190 | | /// </summary> |
| | | 191 | | internal void MarkScanning(DateTime startedUtc, TimeSpan startupDelay, TimeSpan interval) |
| | 2421 | 192 | | => _activation = new WatchdogActivation(Scanning: true, IdleReason: null, startedUtc, startupDelay, interval); |
| | | 193 | | |
| | | 194 | | /// <summary> |
| | | 195 | | /// Marks this host's watchdog as deliberately idle, so the health check can attest "this host |
| | | 196 | | /// does not scan" (the documented multi-host pattern) instead of "no scan yet". |
| | | 197 | | /// </summary> |
| | | 198 | | internal void MarkIdle(string reason) |
| | 38 | 199 | | => _activation = new WatchdogActivation(Scanning: false, reason, StartedUtc: null, default, default); |
| | | 200 | | |
| | | 201 | | /// <summary>How the watchdog resolved its startup guards. Single writer: the watchdog.</summary> |
| | 2459 | 202 | | internal sealed record WatchdogActivation( |
| | 34 | 203 | | bool Scanning, |
| | 12 | 204 | | string? IdleReason, |
| | 9 | 205 | | DateTime? StartedUtc, |
| | 13 | 206 | | TimeSpan StartupDelay, |
| | 2472 | 207 | | TimeSpan Interval); |
| | | 208 | | } |
| | | 209 | | |
| | | 210 | | /// <summary>Result of evaluating a snapshot of the persisted recovery state.</summary> |
| | | 211 | | /// <param name="TotalEntries">Recovery registrations observed, deduplicated per correlation id.</param> |
| | | 212 | | /// <param name="EntriesWithActiveWaiter">Observed entries with at least one live subscriber.</param> |
| | | 213 | | /// <param name="StaleEntries">Entries with no live waiter registered longer ago than the staleness threshold.</param> |
| | | 214 | | /// <param name="UnknownAgeEntries">Entries with no live waiter and no registration timestamp — reported separately, nev |
| | | 215 | | /// <param name="UnprobeableEntries"> |
| | | 216 | | /// Entries whose waiter liveness could not be probed (negative |
| | | 217 | | /// <see cref="RecoveryStateObservation.ActiveSubscribers"/>: the probe failed, or no |
| | | 218 | | /// <see cref="IActiveSubscriberProbe"/> is registered). Their staleness is unknown and they are |
| | | 219 | | /// never flagged stale, so the health check degrades rather than letting a probe outage read as |
| | | 220 | | /// a clean pass with zeroed counters. |
| | | 221 | | /// </param> |
| | | 222 | | /// <param name="Truncated"> |
| | | 223 | | /// Whether the scan stopped at <see cref="AsyncResponseWatchdogOptions.MaxScanEntries"/> before |
| | | 224 | | /// exhausting the store — the counts and stale list then describe the buffered subset only, and |
| | | 225 | | /// the health check degrades rather than attesting a staleness verdict it cannot back. |
| | | 226 | | /// </param> |
| | | 227 | | public sealed record AsyncResponseWatchdogReport( |
| | | 228 | | int TotalEntries, |
| | | 229 | | int EntriesWithActiveWaiter, |
| | | 230 | | IReadOnlyList<RecoveryStateObservation> StaleEntries, |
| | | 231 | | int UnknownAgeEntries, |
| | | 232 | | int UnprobeableEntries = 0, |
| | | 233 | | bool Truncated = false) |
| | | 234 | | { |
| | | 235 | | /// <summary> |
| | | 236 | | /// Pure evaluation: an entry is <em>stale</em> when nobody is subscribed to its channel |
| | | 237 | | /// (the waiter died) and it has been registered for longer than <paramref name="staleAfter"/> |
| | | 238 | | /// without any response triggering the lost-subscriber recovery. Entries without a |
| | | 239 | | /// registration timestamp are reported separately as unknown-age. Entries whose liveness could |
| | | 240 | | /// not be probed (negative <see cref="RecoveryStateObservation.ActiveSubscribers"/>) are never |
| | | 241 | | /// flagged stale, to avoid false positives. |
| | | 242 | | /// </summary> |
| | | 243 | | public static AsyncResponseWatchdogReport Evaluate( |
| | | 244 | | IReadOnlyCollection<RecoveryStateObservation> entries, |
| | | 245 | | DateTime utcNow, |
| | | 246 | | TimeSpan staleAfter) |
| | | 247 | | { |
| | | 248 | | // Dedupe keeps the OLDEST registration per correlation id. Sibling registrations share a |
| | | 249 | | // correlation id by design (fan-out waiters, a flow re-attaching after a crash), and the |
| | | 250 | | // scanner contract deliberately promises no ordering — preferring the oldest makes the |
| | | 251 | | // verdict order-independent, so a young sibling can never mask an older stale one. |
| | | 252 | | // Entries without a correlation id cannot be grouped and are classified individually. |
| | | 253 | | // Structures stay lazily allocated: an empty snapshot allocates nothing. |
| | | 254 | | Dictionary<string, RecoveryStateObservation>? byCorrelationId = null; |
| | | 255 | | List<RecoveryStateObservation>? ungrouped = null; |
| | | 256 | | |
| | | 257 | | foreach (var entry in entries) |
| | | 258 | | { |
| | | 259 | | if (string.IsNullOrEmpty(entry.CorrelationId)) |
| | | 260 | | { |
| | | 261 | | (ungrouped ??= []).Add(entry); |
| | | 262 | | continue; |
| | | 263 | | } |
| | | 264 | | |
| | | 265 | | byCorrelationId ??= new Dictionary<string, RecoveryStateObservation>(entries.Count, StringComparer.Ordinal); |
| | | 266 | | if (!byCorrelationId.TryGetValue(entry.CorrelationId, out var kept) || IsOlder(entry, kept)) |
| | | 267 | | byCorrelationId[entry.CorrelationId] = entry; |
| | | 268 | | } |
| | | 269 | | |
| | | 270 | | var totalEntries = 0; |
| | | 271 | | var entriesWithActiveWaiter = 0; |
| | | 272 | | var unknownAgeEntries = 0; |
| | | 273 | | var unprobeableEntries = 0; |
| | | 274 | | List<RecoveryStateObservation>? staleEntries = null; |
| | | 275 | | |
| | | 276 | | if (byCorrelationId is not null) |
| | | 277 | | { |
| | | 278 | | foreach (var entry in byCorrelationId.Values) |
| | | 279 | | Classify(entry, utcNow, staleAfter, ref totalEntries, ref entriesWithActiveWaiter, ref unknownAgeEntries |
| | | 280 | | } |
| | | 281 | | |
| | | 282 | | if (ungrouped is not null) |
| | | 283 | | { |
| | | 284 | | foreach (var entry in ungrouped) |
| | | 285 | | Classify(entry, utcNow, staleAfter, ref totalEntries, ref entriesWithActiveWaiter, ref unknownAgeEntries |
| | | 286 | | } |
| | | 287 | | |
| | | 288 | | return new AsyncResponseWatchdogReport( |
| | | 289 | | totalEntries, |
| | | 290 | | entriesWithActiveWaiter, |
| | | 291 | | staleEntries ?? [], |
| | | 292 | | unknownAgeEntries, |
| | | 293 | | unprobeableEntries); |
| | | 294 | | } |
| | | 295 | | |
| | | 296 | | /// <summary>Prefers the entry with the oldest known registration; a known age beats an unknown one.</summary> |
| | | 297 | | internal static bool IsOlder(RecoveryStateObservation candidate, RecoveryStateObservation kept) |
| | | 298 | | => candidate.RegisteredAtUtc is { } candidateRegistered |
| | | 299 | | && (kept.RegisteredAtUtc is not { } keptRegistered || candidateRegistered < keptRegistered); |
| | | 300 | | |
| | | 301 | | private static void Classify( |
| | | 302 | | RecoveryStateObservation entry, |
| | | 303 | | DateTime utcNow, |
| | | 304 | | TimeSpan staleAfter, |
| | | 305 | | ref int totalEntries, |
| | | 306 | | ref int entriesWithActiveWaiter, |
| | | 307 | | ref int unknownAgeEntries, |
| | | 308 | | ref int unprobeableEntries, |
| | | 309 | | ref List<RecoveryStateObservation>? staleEntries) |
| | | 310 | | { |
| | | 311 | | totalEntries++; |
| | | 312 | | var activeSubscribers = entry.ActiveSubscribers; |
| | | 313 | | if (activeSubscribers > 0) |
| | | 314 | | { |
| | | 315 | | entriesWithActiveWaiter++; |
| | | 316 | | return; |
| | | 317 | | } |
| | | 318 | | |
| | | 319 | | // Negative liveness means it could not be probed; never flag those as stale, but count |
| | | 320 | | // them — dropped from every bucket, a probe outage would read as a clean pass. |
| | | 321 | | if (activeSubscribers != 0) |
| | | 322 | | { |
| | | 323 | | unprobeableEntries++; |
| | | 324 | | return; |
| | | 325 | | } |
| | | 326 | | |
| | | 327 | | if (entry.RegisteredAtUtc is not { } registeredAtUtc) |
| | | 328 | | { |
| | | 329 | | unknownAgeEntries++; |
| | | 330 | | return; |
| | | 331 | | } |
| | | 332 | | |
| | | 333 | | if (utcNow - registeredAtUtc >= staleAfter) |
| | | 334 | | (staleEntries ??= []).Add(entry); |
| | | 335 | | } |
| | | 336 | | } |
| | | 337 | | |
| | | 338 | | /// <summary> |
| | | 339 | | /// Periodic, report-only scanner of the persisted async-response recovery state. It is part of the |
| | | 340 | | /// engine and runs by default for whatever channel is registered: it enumerates recovery entries |
| | | 341 | | /// through <see cref="IRecoveryStateScanner"/> and checks waiter liveness through |
| | | 342 | | /// <see cref="IActiveSubscriberProbe"/>, so it is independent of any specific store or broker. |
| | | 343 | | /// <para> |
| | | 344 | | /// Every recovery entry represents an outstanding wait registration. A healthy entry either has a |
| | | 345 | | /// live subscriber (the waiter is awaiting in some process) or is young — armed recovery state for |
| | | 346 | | /// a response that has not arrived yet. An entry that is <em>old</em> and has <em>no subscriber</em> |
| | | 347 | | /// means the waiter died and nothing (response, resume, retry) has touched the flow since: the |
| | | 348 | | /// precursor of an operation stuck "in progress". The watchdog logs a warning per such entry and |
| | | 349 | | /// publishes a summary snapshot for the health check. It deliberately performs no remediation — |
| | | 350 | | /// recovery belongs to the lost-subscriber dispatcher and the flows' own retry paths. |
| | | 351 | | /// </para> |
| | | 352 | | /// </summary> |
| | | 353 | | internal sealed class AsyncResponseWatchdog : BackgroundService |
| | | 354 | | { |
| | | 355 | | private readonly IRecoveryStateScanner? _scanner; |
| | | 356 | | private readonly IActiveSubscriberProbe? _subscriberProbe; |
| | | 357 | | private readonly AsyncResponseWatchdogState _state; |
| | | 358 | | private readonly AsyncResponseWatchdogOptions _options; |
| | | 359 | | private readonly TimeProvider _timeProvider; |
| | | 360 | | private readonly ILogger<AsyncResponseWatchdog> _logger; |
| | | 361 | | |
| | | 362 | | /// <summary>Creates the background recovery watchdog.</summary> |
| | | 363 | | public AsyncResponseWatchdog( |
| | | 364 | | IEnumerable<IRecoveryStateScanner> scanners, |
| | | 365 | | IEnumerable<IActiveSubscriberProbe> subscriberProbes, |
| | | 366 | | AsyncResponseWatchdogState state, |
| | | 367 | | IOptions<AsyncResponseOptions> options, |
| | | 368 | | ILogger<AsyncResponseWatchdog> logger, |
| | | 369 | | TimeProvider? timeProvider = null) |
| | | 370 | | { |
| | | 371 | | _scanner = scanners.FirstOrDefault(); |
| | | 372 | | _subscriberProbe = subscriberProbes.FirstOrDefault(); |
| | | 373 | | _state = state; |
| | | 374 | | _options = options.Value.Watchdog; |
| | | 375 | | _options.Validate(); |
| | | 376 | | _timeProvider = timeProvider ?? TimeProvider.System; |
| | | 377 | | _logger = logger; |
| | | 378 | | } |
| | | 379 | | |
| | | 380 | | /// <summary> |
| | | 381 | | /// A base wait plus a random offset up to <see cref="AsyncResponseWatchdogOptions.ResolvedJitter"/>, |
| | | 382 | | /// so replicas that started together do not stay in step. Drawn per wait rather than once per |
| | | 383 | | /// process: a single per-process offset keeps the SPACING identical, so two replicas that |
| | | 384 | | /// happen to draw similar offsets collide on every scan thereafter instead of just the first. |
| | | 385 | | /// </summary> |
| | | 386 | | private TimeSpan NextWait(TimeSpan baseDelay) |
| | | 387 | | { |
| | | 388 | | var jitter = _options.ResolvedJitter; |
| | | 389 | | return jitter <= TimeSpan.Zero |
| | | 390 | | ? baseDelay |
| | | 391 | | : baseDelay + TimeSpan.FromTicks(Random.Shared.NextInt64(jitter.Ticks + 1)); |
| | | 392 | | } |
| | | 393 | | |
| | | 394 | | /// <inheritdoc /> |
| | | 395 | | protected override async Task ExecuteAsync(CancellationToken stoppingToken) |
| | | 396 | | { |
| | | 397 | | if (!_options.Enabled) |
| | | 398 | | { |
| | | 399 | | _state.MarkIdle($"disabled via {nameof(AsyncResponseOptions)}.{nameof(AsyncResponseOptions.Watchdog)}.{nameo |
| | | 400 | | _logger.LogInformation("Recovery watchdog disabled via options; not scanning."); |
| | | 401 | | return; |
| | | 402 | | } |
| | | 403 | | |
| | | 404 | | if (_scanner is null) |
| | | 405 | | { |
| | | 406 | | _state.MarkIdle($"no {nameof(IRecoveryStateScanner)} registered (the configured channel does not support sca |
| | | 407 | | _logger.LogInformation("Recovery watchdog idle: no IRecoveryStateScanner registered (the configured channel |
| | | 408 | | return; |
| | | 409 | | } |
| | | 410 | | |
| | | 411 | | // Only a watchdog that actually scans may take over the process-wide gauge holder: a |
| | | 412 | | // disabled or scanner-less host (the documented multi-host pattern) would otherwise |
| | | 413 | | // permanently zero the gauges for the host that does scan. |
| | | 414 | | _state.MarkScanning(_timeProvider.GetUtcNow().UtcDateTime, _options.StartupDelay, _options.Interval); |
| | | 415 | | AsyncResponseDiagnostics.EnsureWatchdogGauges(_state); |
| | | 416 | | |
| | | 417 | | _logger.LogInformation("Recovery watchdog started. Interval: {Interval}, stale threshold: {StaleAfter}.", _optio |
| | | 418 | | |
| | | 419 | | try |
| | | 420 | | { |
| | | 421 | | await Task.Delay(NextWait(_options.StartupDelay), _timeProvider, stoppingToken).ConfigureAwait(false); |
| | | 422 | | |
| | | 423 | | while (!stoppingToken.IsCancellationRequested) |
| | | 424 | | { |
| | | 425 | | try |
| | | 426 | | { |
| | | 427 | | var report = await ScanOnceAsync(stoppingToken).ConfigureAwait(false); |
| | | 428 | | _state.Publish(new AsyncResponseWatchdogSnapshot(_timeProvider.GetUtcNow().UtcDateTime, _options.Int |
| | | 429 | | } |
| | | 430 | | catch (OperationCanceledException) when (stoppingToken.IsCancellationRequested) |
| | | 431 | | { |
| | | 432 | | break; |
| | | 433 | | } |
| | | 434 | | catch (Exception ex) |
| | | 435 | | { |
| | | 436 | | _logger.LogError(ex, "Recovery watchdog scan failed; next attempt in {Interval}.", _options.Interval |
| | | 437 | | _state.Publish(new AsyncResponseWatchdogSnapshot(_timeProvider.GetUtcNow().UtcDateTime, _options.Int |
| | | 438 | | } |
| | | 439 | | |
| | | 440 | | await Task.Delay(NextWait(_options.Interval), _timeProvider, stoppingToken).ConfigureAwait(false); |
| | | 441 | | } |
| | | 442 | | } |
| | | 443 | | catch (OperationCanceledException) |
| | | 444 | | { |
| | | 445 | | // Host shutdown. |
| | | 446 | | } |
| | | 447 | | } |
| | | 448 | | |
| | | 449 | | /// <inheritdoc /> |
| | | 450 | | public override async Task StopAsync(CancellationToken cancellationToken) |
| | | 451 | | { |
| | | 452 | | try |
| | | 453 | | { |
| | | 454 | | await base.StopAsync(cancellationToken).ConfigureAwait(false); |
| | | 455 | | } |
| | | 456 | | finally |
| | | 457 | | { |
| | | 458 | | AsyncResponseDiagnostics.ReleaseWatchdogGauges(_state); |
| | | 459 | | } |
| | | 460 | | } |
| | | 461 | | |
| | | 462 | | /// <inheritdoc /> |
| | | 463 | | public override void Dispose() |
| | | 464 | | { |
| | | 465 | | // Covers hosts torn down without a graceful StopAsync. The release is identity-conditional |
| | | 466 | | // and idempotent, so the double call on the graceful path is harmless. |
| | | 467 | | AsyncResponseDiagnostics.ReleaseWatchdogGauges(_state); |
| | | 468 | | base.Dispose(); |
| | | 469 | | } |
| | | 470 | | |
| | | 471 | | private async Task<AsyncResponseWatchdogReport> ScanOnceAsync(CancellationToken cancellationToken) |
| | | 472 | | { |
| | | 473 | | using var activity = AsyncResponseDiagnostics.StartActivity("asyncresponse.watchdog.scan"); |
| | | 474 | | |
| | | 475 | | try |
| | | 476 | | { |
| | | 477 | | // Phase 1 — stream the scan and dedupe, keeping the OLDEST registration per |
| | | 478 | | // correlation id (see Evaluate for why oldest). Only the fields the classifier needs |
| | | 479 | | // are buffered, not whole recovery states. Buffering before probing also lets the |
| | | 480 | | // scanner's enumeration (a long-lived reader connection on the relational stores) |
| | | 481 | | // finish before the per-id probe connections open. |
| | | 482 | | Dictionary<string, (DateTime? RegisteredAtUtc, string? PayloadTypeFullName)>? byCorrelationId = null; |
| | | 483 | | List<(string? CorrelationId, DateTime? RegisteredAtUtc, string? PayloadTypeFullName)>? ungrouped = null; |
| | | 484 | | var truncated = false; |
| | | 485 | | int BufferedCount() => (byCorrelationId?.Count ?? 0) + (ungrouped?.Count ?? 0); |
| | | 486 | | |
| | | 487 | | await foreach (var entry in _scanner!.ScanAsync(cancellationToken).ConfigureAwait(false)) |
| | | 488 | | { |
| | | 489 | | if (entry is null) |
| | | 490 | | continue; |
| | | 491 | | |
| | | 492 | | if (string.IsNullOrEmpty(entry.CorrelationId)) |
| | | 493 | | { |
| | | 494 | | // The cap gates growth only — replacing an already-buffered correlation id |
| | | 495 | | // with an older sibling costs nothing, so oldest-wins keeps working at the cap. |
| | | 496 | | if (BufferedCount() >= _options.MaxScanEntries) |
| | | 497 | | { |
| | | 498 | | truncated = true; |
| | | 499 | | break; |
| | | 500 | | } |
| | | 501 | | |
| | | 502 | | (ungrouped ??= []).Add((entry.CorrelationId, entry.RegisteredAtUtc, entry.PayloadTypeFullName)); |
| | | 503 | | continue; |
| | | 504 | | } |
| | | 505 | | |
| | | 506 | | byCorrelationId ??= new Dictionary<string, (DateTime?, string?)>(StringComparer.Ordinal); |
| | | 507 | | if (byCorrelationId.TryGetValue(entry.CorrelationId, out var kept)) |
| | | 508 | | { |
| | | 509 | | if (entry.RegisteredAtUtc is { } candidate && (kept.RegisteredAtUtc is not { } existing || candidate |
| | | 510 | | byCorrelationId[entry.CorrelationId] = (entry.RegisteredAtUtc, entry.PayloadTypeFullName); |
| | | 511 | | } |
| | | 512 | | else |
| | | 513 | | { |
| | | 514 | | if (BufferedCount() >= _options.MaxScanEntries) |
| | | 515 | | { |
| | | 516 | | truncated = true; |
| | | 517 | | break; |
| | | 518 | | } |
| | | 519 | | |
| | | 520 | | byCorrelationId[entry.CorrelationId] = (entry.RegisteredAtUtc, entry.PayloadTypeFullName); |
| | | 521 | | } |
| | | 522 | | } |
| | | 523 | | |
| | | 524 | | // Phase 2 — one liveness probe per unique correlation id, fanned out with bounded |
| | | 525 | | // concurrency (each probe is its own channel round trip; strictly sequential awaits |
| | | 526 | | // would serialize up to MaxScanEntries of them per scan), then the same pure |
| | | 527 | | // classifier the report type exposes publicly, so this scan and Evaluate (the tested |
| | | 528 | | // and benchmarked surface) can never drift apart again. |
| | | 529 | | var pending = new List<(string? CorrelationId, DateTime? RegisteredAtUtc, string? PayloadTypeFullName)>(Buff |
| | | 530 | | |
| | | 531 | | if (byCorrelationId is not null) |
| | | 532 | | { |
| | | 533 | | foreach (var (correlationId, entry) in byCorrelationId) |
| | | 534 | | pending.Add((correlationId, entry.RegisteredAtUtc, entry.PayloadTypeFullName)); |
| | | 535 | | } |
| | | 536 | | |
| | | 537 | | if (ungrouped is not null) |
| | | 538 | | pending.AddRange(ungrouped); |
| | | 539 | | |
| | | 540 | | // Per-slot writes keep the result independent of probe completion order. A probe |
| | | 541 | | // canceled by shutdown still aborts the whole fan-out: CountActiveSubscribersAsync |
| | | 542 | | // rethrows, and ForEachAsync cancels its siblings and surfaces the cancellation. |
| | | 543 | | var observations = new RecoveryStateObservation[pending.Count]; |
| | | 544 | | await Parallel.ForEachAsync( |
| | | 545 | | Enumerable.Range(0, pending.Count), |
| | | 546 | | new ParallelOptions { MaxDegreeOfParallelism = _options.ProbeConcurrency, CancellationToken = cancellati |
| | | 547 | | async (index, ct) => |
| | | 548 | | { |
| | | 549 | | var (correlationId, registeredAtUtc, payloadTypeFullName) = pending[index]; |
| | | 550 | | observations[index] = new RecoveryStateObservation( |
| | | 551 | | correlationId, |
| | | 552 | | registeredAtUtc, |
| | | 553 | | await CountActiveSubscribersAsync(correlationId, ct).ConfigureAwait(false), |
| | | 554 | | payloadTypeFullName); |
| | | 555 | | }).ConfigureAwait(false); |
| | | 556 | | |
| | | 557 | | var report = AsyncResponseWatchdogReport.Evaluate(observations, _timeProvider.GetUtcNow().UtcDateTime, _opti |
| | | 558 | | |
| | | 559 | | if (truncated) |
| | | 560 | | { |
| | | 561 | | // Carried on the report itself, not just telemetry: the health check and gauges |
| | | 562 | | // read the report, and a silently truncated scan would otherwise attest a |
| | | 563 | | // staleness verdict it never actually computed. |
| | | 564 | | report = report with { Truncated = true }; |
| | | 565 | | activity?.SetTag("asyncresponse.watchdog.truncated", true); |
| | | 566 | | _logger.LogWarning( |
| | | 567 | | "Recovery watchdog scan stopped at the {MaxScanEntries}-entry buffer cap; staleness is reported for |
| | | 568 | | _options.MaxScanEntries); |
| | | 569 | | } |
| | | 570 | | |
| | | 571 | | activity?.SetTag("asyncresponse.watchdog.total_entries", report.TotalEntries); |
| | | 572 | | activity?.SetTag("asyncresponse.watchdog.active_waiters", report.EntriesWithActiveWaiter); |
| | | 573 | | activity?.SetTag("asyncresponse.watchdog.stale_entries", report.StaleEntries.Count); |
| | | 574 | | activity?.SetTag("asyncresponse.watchdog.unknown_age_entries", report.UnknownAgeEntries); |
| | | 575 | | activity?.SetTag("asyncresponse.watchdog.unprobeable_entries", report.UnprobeableEntries); |
| | | 576 | | |
| | | 577 | | _logger.LogInformation("Recovery watchdog scan complete. Outstanding registrations: {Total}, with live waite |
| | | 578 | | |
| | | 579 | | foreach (var stale in report.StaleEntries) |
| | | 580 | | { |
| | | 581 | | _logger.LogWarning("Stale async-response recovery state — correlationId {CorrelationId}, payload type {P |
| | | 582 | | } |
| | | 583 | | |
| | | 584 | | return report; |
| | | 585 | | } |
| | | 586 | | catch (Exception ex) |
| | | 587 | | { |
| | | 588 | | AsyncResponseDiagnostics.SetError(activity, ex); |
| | | 589 | | throw; |
| | | 590 | | } |
| | | 591 | | } |
| | | 592 | | |
| | | 593 | | /// <summary> |
| | | 594 | | /// Liveness for one entry. Returns <c>-1</c> (unknown) when there is no probe or no correlation |
| | | 595 | | /// id; the report treats unknown liveness as "not stale" so it never raises a false alarm. |
| | | 596 | | /// </summary> |
| | | 597 | | private async ValueTask<long> CountActiveSubscribersAsync(string? correlationId, CancellationToken cancellationToken |
| | | 598 | | { |
| | | 599 | | if (_subscriberProbe is null || string.IsNullOrWhiteSpace(correlationId)) |
| | | 600 | | return -1; |
| | | 601 | | |
| | | 602 | | try |
| | | 603 | | { |
| | | 604 | | return await _subscriberProbe.CountActiveSubscribersAsync(correlationId, cancellationToken).ConfigureAwait(f |
| | | 605 | | } |
| | | 606 | | catch (OperationCanceledException) when (cancellationToken.IsCancellationRequested) |
| | | 607 | | { |
| | | 608 | | // Shutdown must abort the scan here, not degrade to -1: swallowed, a canceled probe |
| | | 609 | | // would let the loop grind through every remaining entry (throw/catch per id, delaying |
| | | 610 | | // shutdown) and then publish a snapshot attesting a completed scan whose liveness was |
| | | 611 | | // never probed. ExecuteAsync catches this as its stop signal. |
| | | 612 | | throw; |
| | | 613 | | } |
| | | 614 | | catch (Exception ex) |
| | | 615 | | { |
| | | 616 | | _logger.LogDebug(ex, "Recovery watchdog failed to probe subscribers for correlationId {CorrelationId}.", cor |
| | | 617 | | return -1; |
| | | 618 | | } |
| | | 619 | | } |
| | | 620 | | } |