indicator-stream

StreamHub indicator development

Provider selection

Self-rooted source hubs (those that originate a stream rather than transform another hub's output) take an inert sentinel provider so the base-class constructor has something to subscribe to; the sentinel rejects subscriptions and carries no cache.

Aggregator / quantizer hubs

Hubs that bucket small bars (or raw ticks) into larger time periods derive from BarProvider<TIn, IBar>. Conventions:

• Constructors accept a BarInterval enum and a custom TimeSpan overload; the enum overload throws for month-or-longer periods (use TimeSpan instead) since calendar arithmetic is not a fixed TimeSpan.
• Take an optional fillGaps flag. Default false (silent buckets are simply omitted from the output stream); true synthesizes zero-volume bars whose Open/High/Low/Close all carry forward the prior bar's close through the silent period.
• Round the input timestamp down to the current bucket on every OnAdd, then either update the current bar in place or emit a new bucket.
• Override Rebuild(DateTime) to align the requested rebuild timestamp to the bucket boundary before delegating to base — an upstream rebuild whose timestamp is mid-bucket must clear the in-cache partial bar, not duplicate it.
• Implement RollbackState(int) to reset the in-flight bar state and prune any per-input tracker (e.g. duplicate-detection map) past the rollback point.

Aggregator hubs ship full StreamHub semantics: late-arriving inputs whose timestamp lands in an already-emitted bucket trigger a Rebuild of that bucket; downstream observers see the corrected sequence.

Performance targets

Use the project's performance-analysis document as the source of truth for measured overhead bands; the categorical targets below are guidance, not contracts.

The "framework floor" is the per-tick overhead inherent to the observer pattern, cache management, and read-only collection wrappers. Simple stateless indicators routinely measure 6–11x against Series while still achieving tens of thousands of bars per second; this is acceptable. Optimization effort should target indicator-specific algorithmic issues, not the framework floor.

Forbid O(n²) recalculation — rebuild entire history on each tick:

// WRONG
for (int k = 0; k <= i; k++) { subset.Add(cache[k]); }
var result = subset.ToIndicator();

O(1) incremental update:

// CORRECT
_avgGain = ((_avgGain * (period - 1)) + gain) / period;

Use RollingWindowMax/Min utilities instead of O(n) linear scans.

Thread safety contract

StreamHub mutating operations (Add, Rebuild, RemoveRange, RemoveAt) hold a private monitor for the duration of cache mutation, and observer notification happens inside the lock so subscribers cannot desynchronize. Subclasses must not release the lock before notifying observers.

The base class also carries a rebuilding flag that suppresses self-recursive Rebuild while replaying provider items through OnAdd. Observer cascading is still allowed and desired. Subclass code must not bypass this flag.

The public Results surface is a live read-only view over the cache, not an immutable snapshot — and .ToList()/.ToArray() on it enumerate that live view without the lock, so they can still throw or tear under a concurrent writer. A consumer on a different thread must call Snapshot() (an atomic, immutable copy taken under the hub's CacheLock) instead.

RollbackState pattern

Override when maintaining stateful fields. The base class computes restoreIndex via IndexBefore before calling this method. restoreIndex is the last ProviderCache index to preserve, or -1 to reset everything.

protected override void RollbackState(int restoreIndex)
{
    _window.Clear();
    if (restoreIndex < 0) return;
    int startIdx = Math.Max(0, restoreIndex + 1 - LookbackPeriods);
    for (int p = startIdx; p <= restoreIndex; p++)
        _window.Add(ProviderCache[p].Value);
}

Replay up to restoreIndex (inclusive). The item at the rollback timestamp is recalculated via normal processing.

Testing requirements

• Inherit StreamHubTestBase
• Abstract method (compile error if missing): ToStringOverride_ReturnsExpectedName()
• Implement ONE observer interface:
  • ITestChainObserver (most indicators — chain input): inherits ITestBarObserver, adds ChainObserver_ChainedProvider_MatchesSeriesExactly()
  • ITestBarObserver (direct bar input only): BarObserver_WithWarmupLateArrivalAndRemoval_MatchesSeriesExactly(), WithCachePruning_MatchesSeriesExactly()
• If hub acts as chain provider, also implement ITestChainProvider: ChainProvider_MatchesSeriesExactly()

Required implementation

• Source code: src/**/{IndicatorName}Hub.cs file exists
  • Uses appropriate provider base (ChainHub or BarProvider)
  • Validates parameters in constructor; calls Reinitialize() as needed
  • Implements O(1) state updates; avoids O(n²) recalculation
  • Overrides RollbackState() when maintaining stateful fields
  • Overrides ToString() with concise hub name
• Unit testing: tests/Library/Indicators/**/{IndicatorName}HubTests.cs exists
  • Inherits StreamHubTestBase with correct test interfaces
  • Comprehensive rollback validation present
  • Verifies Series parity
• Catalog registration: Registered in Catalog.Listings.cs
• Performance benchmark: Add to tools/performance/Perf.Stream.cs
• Public documentation: Update docs/indicators/{IndicatorName}.md
• Regression tests: Add to tests/Library/Indicators/**/{IndicatorName}RegressionTests.cs
• Migration guide: Update docs/migration/v3.md for notable and breaking changes from v2

References

• Provider selection
• Rollback patterns
• Performance patterns
• Compound hubs

Constraints

• O(n²) recalculation is forbidden; all updates must be O(1)
• RollbackState(int restoreIndex) receives the last index to preserve (-1 = reset all); replay is inclusive of restoreIndex, exclusive of the rollback timestamp
• Series parity required: results must be numerically identical to StaticSeries