// the find
sryza/spark-timeseries
A library for time series analysis on Apache Spark
A Scala/Java/Python library for distributed time series analysis on Spark, covering ARIMA, GARCH, EWMA, Holt-Winters, and statistical tests like Augmented Dickey-Fuller. It's the closest thing the Spark ecosystem had to a pandas-timeseries equivalent before the author formally stepped back in 2018. Last commit was 2020; treat it as archived.
The TimeSeriesRDD abstraction is well-designed — it aligns multiple time series on a shared DateTimeIndex so windowed ops and resampling don't require per-series bookkeeping. Statistical model coverage is decent: ARIMA, ARIMAX, GARCH, and AR variants with Python bindings that delegate to the JVM, keeping a single implementation. Uses Breeze for BLAS-accelerated linear algebra rather than reinventing it. The test suite validates models against R-generated reference data, which is the right way to verify statistical correctness.
Abandoned — the author's own README says so, and the last real activity predates Spark 3.x, meaning you'll fight version incompatibilities immediately on any modern Spark cluster. No streaming support; everything is batch RDD-based, which misses the obvious use case of live sensor or market data. The Python layer is a thin JVM bridge with no native path, so debugging PySpark errors means chasing through Java stack traces. No support for irregular or multi-frequency time series in the same RDD — you need identical DateTimeIndex alignment across all series, which breaks on real-world financial or IoT data.