Module com.darkcollective.relix.optimizer
Package com.darkcollective.relix.optimizer.internal

Class SelectionIntoTimeSeriesPass

java.lang.Object
com.darkcollective.relix.optimizer.internal.SelectionIntoTimeSeriesPass
public final class SelectionIntoTimeSeriesPass extends Object
Partition-pruning pass (ADR-0020) for the two time-series operators (SESSION-001 / DOWNSAMPLE-001). 
   σ k = c (SESSIONIZE ts GAP … PER k AS s (R))  →  SESSIONIZE … PER k AS s (σ k = c (R))
   σ k = c (DOWNSAMPLE ts BY '5m' USING AVG PER k (R))
                                                 →  DOWNSAMPLE … PER k (σ k = c (R))

Both compute independently per partition — a session boundary never spans a PER key, and a bucket is keyed by (PER keys, bucket) — which is the identical soundness argument ADR-0020 already accepted for WINDOW, TOP and OPTIMIZE. Both are also MaterializationMode.BAG: they buffer and process every partition to produce the one the query asked for, so the saving is real work, not just rows discarded a little later.

The descriptor and the traversal are PartitionPruning's; this class supplies only the two node shapes.

FOR n ROWS blocks the DOWNSAMPLE push

DownsampleNode's maxRows keeps the n most recent buckets globally, not per group — DownsampleExecutor sorts all output rows by bucket descending and takes the first n. So with two groups and FOR 2 ROWS, filtering afterwards can leave one row while pushing the filter first would produce two. The rule therefore does not fire at all when maxRows is present. (The issue did not flag this; it is a genuine counterexample rather than a conservative choice.)

Out of scope, and noted as a follow-on: a predicate on DOWNSAMPLE's timestamp column. It is a range restriction on the buckets rather than a partition selection — a different and more interesting rewrite, which has to reason about bucket boundaries rather than just about which groups exist.

This class is package-private and stateless; call apply(RelNode, String, SchemaAnnotations, OptimizationContext) as a static method.