ValkeyTimeSeriesPreview

ValkeyTimeSeries is loaded into a Valkey server with --loadmodule. Each time series is a key holding labels (name-value pairs) and a list of chunks: ordered, non-overlapping blocks of samples, each a 64-bit timestamp and a 64-bit float, stored compressed by default. TS.ADD and TS.MADD add samples, creating a series if needed; TS.RANGE and TS.MRANGE read a time range from one series or from every series whose labels match a filter, optionally aggregated into buckets; TS.CREATERULE keeps a downsampled series up to date as samples arrive. A retention period drops old samples, actively in the background rather than lazily on query. The command set is meant to be a superset of RedisTimeSeries, so its clients, such as @valkey/time-series, the node-valkey client’s copy of node-redis’s RedisTimeSeries package, send the same commands.

Its additions go beyond RedisTimeSeries: compaction rules can chain (minutes to hours to days), filters accept Prometheus series selectors such as latency{region=~"us-west-*"}, TS.JOIN joins two series on timestamp, including as-of joins within a tolerance, TS.OUTLIERS flags anomalies with z-score, IQR, MAD or random cut forest, and metadata commands list label names, values and cardinality. Label lookups use an inverted index: an adaptive radix tree from each label pair to a Roaring bitmap of series IDs. Queries and batched inserts run in parallel across series and chunks, and identical label strings are stored once. It works in standalone or cluster mode, where related series should share a hash tag so a query stays on one node.