TimeDB stores overlapping forecast revisions, auditable corrections, and "time-of-knowledge" history as plain rows in ClickHouse. Every value carries a three-dimensional timestamp so you can replay exactly what was known at any past instant.

Traditional time-series stores assume one immutable value per timestamp. TimeDB is built for the messy reality of forecasts that get revised, observations that get corrected, and backtests that need strict point-in-time audits.

🧊 The 3D Temporal Data Model

At the heart of TimeDB is its three-dimensional approach to time. We track not just when data is valid, but when it became known and when it was altered.

Audit & Metadata: Every row also carries changed_by and annotation text fields so corrections leave a readable trail instead of silent overwrites.

✨ Why Choose TimeDB?

• 📊 Forecast Revisions: Store overlapping forecasts side-by-side — every knowledge_time is preserved.
• 🔄 Auditable Updates: Corrections are new rows with a fresh change_time; reads collapse them into the latest state, with full history available on demand.
• ⏪ True Backtesting: Query historical data as of any point in time ("What did our model know last Monday?"), or use read_relative() for per-window day-ahead cutoffs.
• 🗂️ Retention Tiers: Pick short / medium / long / forever per series; ClickHouse drops whole partitions when TTLs expire.
• 🧹 Idempotent Writes: Opt into write(..., skip_unchanged=True) and rows that only duplicate the latest stored value are dropped before insert — writing the same window repeatedly stops bloating storage. Off by default; full provenance kept unless you ask. unchanged_scope="auto" applies a different comparison key per series in one call, so a frame mixing plain actuals with versioned publications dedupes each correctly.
• 🪶 Stateless & Minimal: One class, two tables, no catalog. Series identity (series_id) is owned by the caller — no naming, units, or labels to keep in sync.

🚀 Quick Start

1. Installation

pip install timedb

Requires Python 3.12+ and a reachable ClickHouse instance. TimeDB reads its connection string from TIMEDB_CH_URL (also picked up from a .env file).

2. Usage Example

import polars as pl
from datetime import datetime, timezone
from timedb import TimeDBClient

td = TimeDBClient()  # reads TIMEDB_CH_URL
td.create()          # creates series_values + run_series tables

# 1. Write a forecast run for series_id=42, issued at 06:00.
kt = datetime(2025, 1, 1, 6, tzinfo=timezone.utc)
df = pl.DataFrame({
    "series_id":  [42] * 24,
    "valid_time": [datetime(2025, 1, 1, h, tzinfo=timezone.utc) for h in range(24)],
    "value":      [100.0 + i * 2 for i in range(24)],
})
td.write(df, retention="medium", knowledge_time=kt)

# 2. A later forecast revision — same valid_time window, higher knowledge_time.
kt2 = datetime(2025, 1, 1, 12, tzinfo=timezone.utc)
td.write(df.with_columns(pl.col("value") + 5), retention="medium", knowledge_time=kt2)

# 3. Latest value per valid_time (the second run wins).
latest = td.read(series_ids=[42])

# 4. Full forecast history — one row per (knowledge_time, valid_time).
history = td.read(series_ids=[42], include_knowledge_time=True)

Required write columns are series_id, valid_time, value. Everything else (change_time, run_id, changed_by, annotation, valid_time_end) is optional and stamped with safe defaults per batch. All timestamp columns must be timezone-aware.

🤝 The Full Stack — TimeDB + EnergyDB

TimeDB stores rows keyed by integer series_id and nothing else. That's enough when you're managing identity in your own application — but for the energy domain, we ship the full stack.

EnergyDB adds, on top of the same ClickHouse store (plus a thin PostgreSQL catalog):

• 🌳 Typed asset trees: Portfolio → Site → WindTurbine / PVArray / Battery, etc. — every asset class from EnergyDataModel.
• 🔗 Grid edges: Line, transformer, pipe — connect any two nodes, attach their own time series.
• 🧭 Fluent path scopes: client.get_node("my-portfolio", "Offshore-1", "T01").read(name="power", data_type="actual") resolves to one indexed SQL query.
• ⚖️ Per-series canonical units: declare MW once; pint converts on every read and write via a unit= kwarg.
• 🧬 Run/workflow provenance: workflow_id, model_name, run_start_time, etc. attached at write time.
• 📋 Diffable structural changes: dry_run=True previews every rename, move, delete, or insert as a TreeDiff before you commit.

Use TimeDB for the raw storage primitive. Use EnergyDB for an asset-aware catalog of energy portfolios.

🧪 Try it in Google Colab

Want to try TimeDB without a local setup? Open our Quickstart in Colab — the first cell automatically installs ClickHouse inside the VM.

Note: Data persists only within the active Colab session. Additional notebooks are available in the examples/ directory.

📚 Documentation & Resources

• 📖 Official Documentation
• ⚙️ Installation Guide
• 🐍 SDK Guide
• 🌐 Reference
• 💡 Examples & Notebooks

🤝 Contributing

Contributions are welcome! If you're interested in improving TimeDB, please see our Development Guide for local setup instructions.