Documentation
¶
Overview ¶
Package featstream is the one spec-level streaming verdict for the built-in FEAT_* operators, shared by the engine's streaming gate (internal/processing/feature IsStreamableWithExt / BuildStreamingWithExt) and predict's streamable reasons (internal/descriptor computeStreamable), so the two cannot disagree.
The streaming feature drive runs every feature's pre-pass over the RAW record (pass 1) before any feature has emitted (pass 2). A global-pass operator whose pre-pass reads a column an earlier feature derives sees it absent there and would fold nothing — every value null or every quantile cut at zero — while the buffered path, which applies features in order, computes it. Such a spec routes buffered:
- FEAT_TARGET_ENCODE: its field, target or any name params.fit_on reads (internal/targetencode.Streamable);
- FEAT_BUCKETIZE in quantile mode: its field.
The other built-ins either read only stored levels (FEAT_FREQUENCY_ENCODE and FEAT_TRAIN_TEST_SPLIT read categorical / packed_bool levels, which a derived numeric column never carries — null on both arms alike) or are row-local (their pre-pass is a no-op and pass 2 emits in order).
Leaf: the public encoding / types packages + targetencode + paramread (a Probe: the operator itself decodes and records its params).
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func Streamable ¶
Streamable reports whether feat can run on the streaming feature drive over schema; when it cannot, reason says why (predict's streamable_reasons entry). Params that do not decode leave the verdict to the operator's own refusal (true here). A nil schema stores no field.
Types ¶
This section is empty.
Source Files