TickerFlow

TickerFlow is a Python backend that turns local market CSVs into validated, queryable API data.

Open the live demo → (synthetic OHLCV fixtures)

Start reviewing here

Problem: make local OHLCV data quality decisions visible before data is queryable.

• Workflow: CSV → canonical UTC frame → validation report → partitioned Parquet → half-open queries → hourly/daily time bars → FastAPI demo.
• Hard decision: quarantine invalid rows without silent repair. Quality policy and the invalid-OHLC regression.
• Validation: dirty fixture report tests and ingestion/storage/query regressions.
• Quick run: uv sync --extra dev && uv run pytest; then follow the complete seed → query demonstration.
• Limits: local filesystem and one writer; identical-key repeat ingestion is idempotent, but does not establish concurrent-write safety. Writes directly replace partitions; no atomic crash recovery or multi-partition transactions are promised. Trades/quotes and tick/volume/dollar bars remain planned.

Use case

Market-data pipelines need deterministic ingestion, explicit schemas, visible quality decisions, durable storage, and stable query boundaries. TickerFlow provides that local workflow for financial time series, from CSV normalization through Parquet storage and feature-ready time bars.

Implemented capabilities

• Ingest local OHLCV CSV files with explicit schemas and configuration.
• Normalize timestamps to UTC and report data-quality issues.
• Store valid OHLCV rows as partitioned Parquet.
• Query datasets by symbol and half-open date range through Python and FastAPI.
• Build hourly and daily time bars with explicit interval boundaries.
• Explore the local workflow through the /demo page.

Planned extensions

• Trade and quote ingestion.
• Tick, volume, and dollar bars.
• Reproducible performance benchmarks on larger datasets.

Technology

• Python 3.12+
• Polars for DataFrame transformations.
• DuckDB for local analytical queries.
• PyArrow/Parquet for storage.
• Pydantic and FastAPI for backend contracts.
• pytest, ruff, and mypy for quality.

Repository philosophy

1. Schemas before pipelines.
2. Deterministic local fixtures before live data.
3. Validation reports before silent cleaning.
4. Small vertical slices over broad unsupported features.

First milestones

1. Bootstrap Python package, CI, linting, typing, and tests (done).
2. Define canonical schemas for OHLCV (done) and trades (planned).
3. Implement local CSV ingestion with validation reports (done).
4. Store normalized data as partitioned Parquet (done).
5. Add query service and FastAPI endpoint (done).
6. Implement time bars (done); volume and dollar bars remain planned.
7. Add benchmarks on larger datasets (planned); quality-report examples ship as the demo cases.

Current vertical slice

TickerFlow currently implements the first OHLCV backend slice:

• Load local OHLCV CSV files with typed ingestion configuration.
• Normalize timestamps to timezone-aware UTC at microsecond precision.
• Validate dirty rows and return a structured quality report instead of silently dropping data.
• Store valid rows as partitioned local Parquet under ohlcv/symbol=<SYMBOL>/date=<YYYY-MM-DD>/data.parquet.
• Query OHLCV rows by symbol and half-open UTC date range [start, end).
• Discover available local datasets and symbols from Parquet partitions.
• Build hourly or daily time bars with explicit half-open boundaries.
• Expose /health, /datasets, /symbols, /ohlcv, and /bars/time through FastAPI.
• Provide a browser market-data demo at /demo.

OHLCV schema assumptions

Input CSV fixtures use these columns:

timestamp,symbol,open,high,low,close,volume

Canonical rows use:

timestamp_utc: datetime[us, UTC]
symbol: uppercase string
open: float
high: float
low: float
close: float
volume: float
source: string

Prices are unadjusted fixture values in arbitrary currency units. OHLC prices must be finite, positive, and consistent with low/high bounds. Volume is a finite non-negative numeric quantity. Corporate actions and live data sources are intentionally out of scope.

Local development

uv sync --extra dev
uv run ruff format .
uv run ruff check .
uv run mypy src tests
uv run pytest

Run the API locally:

uv run uvicorn tickerflow.api.main:app --reload

Open the market-data demo UI:

open http://127.0.0.1:8000/demo

Example query after writing Parquet data into the configured data directory:

curl "http://127.0.0.1:8000/ohlcv?symbol=AAPL&start=2024-01-02T00:00:00Z&end=2024-01-04T00:00:00Z"

Catalog endpoints:

curl "http://127.0.0.1:8000/datasets"
curl "http://127.0.0.1:8000/symbols?dataset=ohlcv"

Time-bar endpoint:

curl "http://127.0.0.1:8000/bars/time?symbol=AAPL&start=2024-01-02T14:00:00Z&end=2024-01-02T16:00:00Z&interval=1h"

By default, the API reads from .tickerflow. Set TICKERFLOW_DATA_DIR to point at another local Parquet root.

The deterministic demo script lives in docs/DEMO.md.

Native homelab deployment

The prepared homelab manifest uses sudo homelab project deploy tickerflow, running the repository quality gate before activating an immutable release; homelab is the author's private server tooling, not part of this repository. Uvicorn is a runtime dependency, retained by production-only uv sync --frozen --no-extra dev --no-editable. The API binds loopback behind Caddy; /health checks readiness and the existing root redirects to /demo. Set TICKERFLOW_DATA_DIR to writable state outside the release (the VPS manifest uses /var/lib/homelab/tickerflow/data). Demo input remains synthetic. The demo is published at https://tickerflow.lucaslebihan.dev/demo.

Market-data investigation demo

The main /demo uses bounded, committed cases through GET /demo/cases and POST /demo/run. Each run keeps original string cells, normalized values and actual row-level validation decisions, writes accepted rows to a request-local Parquet store, queries with DuckDB and traces query rows into hourly/daily bars. Temporary files are cleaned after the request; the shared persistent dataset is not altered. The legacy /demo/seed and data/query APIs remain compatible.

Accepted + quarantined counts reconcile to the same input. Issue counts can overlap. Duplicate pairs are both quarantined. Invalid original values remain inspectable even when normalization produces null/non-finite values; response JSON uses safe nulls for non-finite normalized cells. Cases are synthetic, unadjusted OHLCV, UTC microseconds, currency-unit prices and synthetic volume.

Bars aggregate only the selected query's observations, use half-open UTC intervals and do not fill gaps. Clipped edge intervals are marked. Exports contain actual report/query/bar data and fixture hashes. Repeating a run demonstrates canonical result reproducibility, not byte-identical Parquet files. No upload or arbitrary SQL endpoint is exposed. The production wheel includes fixtures/assets. Set DEMO_BUILD_REVISION to the actual build SHA or accept the honest unknown.

Run browser checks with node tests/browser/demo.cjs, Playwright available and TICKERFLOW_DEMO_URL pointing at a local server. Hand-checked case expectations live in tests/fixtures/demo_expected.json.