Adaptive Sampling Under Extreme Tick Rates

When to Use

Invoke this skill for a high-rate trade-tick stream when a downstream strategy or analytics consumer has a measured processing ceiling and cannot safely keep every input tick. The engine monitors per-symbol arrival rate, emits every 1:N ticks during overload, and aggregates skipped trade volume and price-volume notional into the emitted sample.

The default target of 5,000 ticks/second is an example capacity parameter, not a universal market-data standard. Calibrate it from measured CPU, queue, latency, and downstream correctness budgets.

When NOT to Use

• Do not use sampled output for order-book reconstruction, quote protection, spread/microstructure signals, queue-position models, latency-sensitive execution, or regulatory/compliance records that require the complete feed.
• Do not use it to repair feed gaps, sequence gaps, duplicate messages, or venue recovery state. Detect and reconcile those upstream before sampling.
• Do not drop residual aggregates at shutdown, symbol removal, or feed restart; call flush, flush_all, or reset_symbol(flush=True) and persist the result.
• Do not treat aggregate VWAP as an OHLC bar or assume it preserves the path, timing, or extrema of skipped trades.

Prerequisites

• A trade-tick feed with documented symbol identity, sequence semantics, event timestamps, price units, and volume units.
• A measured per-symbol processing target and an explicit policy for overload, feed restart, sequence reset, and symbol lifecycle.
• Downstream consumers that understand SampledTick.aggregated_tick_count, sampling_factor, and synthetic is_flush records.
• A durable handoff or audit path for emitted samples, validation errors, and flush/recovery events.
• Python 3.10+.

Workflow

1. Set the capacity contract: Construct AdaptiveTickSamplerEngine(target_max_rate_per_sec=...) with a positive integer target. Keep enforce_monotonic_sequence=True unless the feed contract explicitly permits another policy.
2. Validate feed identity upstream: Verify symbol, sequence, timestamp, price, and volume semantics. The engine rejects empty symbols, non-finite values, non-positive prices/volumes, duplicate or decreasing sequences, and backwards event timestamps.
3. Ingest one trade tick at a time: Call ingest_tick(...). Under the target rate it emits a passthrough aggregate; above the target it computes k = ceil(rate / target) with a minimum sampled factor of 2.
4. Consume explicit metadata: Use mode and sampling_factor to identify policy, aggregated_tick_count to identify how many raw trades are represented, and is_flush to distinguish synthetic residual output. Treat aggregated_tick_count > 1 — not mode — as the test for whether price is a VWAP rather than a traded price, because a PASSTHROUGH emission also drains any residual sampled block.
5. Handle errors fail-closed: Route validation exceptions to the feed-integrity path. Do not retry the same malformed or duplicate tick as if it were new data.
6. Manage lifecycle: Call flush(symbol) for an individual stream, flush_all() during shutdown/checkpointing, and reset_symbol(flush=True) when a symbol leaves the stream or its sequence domain resets.
7. Monitor quality: Track input/output rates, sampling factor, aggregate counts, volume/notional reconciliation, validation failures, flush counts, queue latency, and downstream processing latency.

Common Pitfalls

• Discarding skipped volume: Emitting only the selected tick loses traded volume and corrupts aggregate VWAP.
• Calling the result a full feed: A VWAP aggregate preserves volume and notional, not every price path, quote, or order-book event.
• Keying on mode to detect a raw trade: when the rate falls back below target mid-block, the draining emission reports PASSTHROUGH with sampling_factor=1 while still carrying aggregated_tick_count > 1 and a VWAP price no venue printed. Only aggregated_tick_count separates a raw trade from an aggregate.
• Accepting duplicate or out-of-order data: Repeated trades double-count volume; backwards event time corrupts rolling-rate windows.
• Using wall-clock flush timestamps: A flush must remain deterministic in replay; the implementation defaults to the last event timestamp unless an explicit later timestamp is supplied.
• Leaking symbol state: Long-lived feeds must reset inactive symbols after flushing so per-symbol dictionaries do not grow without bound.
• Sampling before feed recovery: Sequence gaps and venue recovery require upstream reconciliation; sampling cannot infer missing trades.

Verification

Run the focused test suite:

python -m unittest discover -s skills/adaptive-sampling-under-extreme-tick-rates/scripts

The tests cover passthrough, the rate == target boundary, overload sampling, volume/notional preservation across a burst, the PASSTHROUGH drain of a partial sampled block, zero timestamps, deterministic flushes, rejection of a flush timestamp preceding the last accepted tick, sorted flush_all, reset_symbol with and without a flush, duplicate/out-of-order rejection, the optional non-monotonic sequence policy, malformed tick values, invalid targets, symbol isolation, state-neutral rejection of an overflowing aggregate, and concurrent multi-threaded ingestion of a shared symbol. Production sign-off additionally requires replaying calibrated bursts and verifying downstream reconciliation against raw input totals.

Related Skills

• backpressure-drop-degrade-policy
• tick-buffering-burst-handling
• adaptive-batch-size-tuning-under-load