kubernetes-sigs / inference-perf / 38476258500

86%

push

github

Add `numeric/expression` (#563)

Adds `utils/numeric/expression`: an `Expression` type that turns a
numeric config knob written as a string into an object parsed and
validated once at construction, then sampled cheaply many times, plus
`Predicate`, its boolean counterpart for stop conditions.

## Motivation

Numeric config knobs today are fixed scalars or structured
`Distribution` configs. There is no compact way to express time-varying
load shapes (ramps, sinusoids, bursts) or ad-hoc randomness, and every
new shape means new config surface. One string grammar covers all three
and their combinations:

- constants: `"3.5"`
- functions of stage time: `"10 + 5*sin(2*pi*t/60)"`
- draws from any `sympy.stats` distribution: `"Normal(512, 200)"`, `"100
+ 50*Normal(0, 1)"`

Each call site decides which ingredients are permitted (`allow_time`,
`allow_random`) and may bound the range, so a request-rate knob can
forbid negatives. `Expression` complements `numeric.distribution` rather
than replacing it.

## Change structure

`inference_perf/utils/numeric/expression/expression.py`:

- **Distribution discovery**: every `sympy.stats` constructor is found
at import and bound without a name argument; unique generated names keep
repeated draws independent.
- **`Expression.__init__`**: parse, reject unknown functions, enforce
`allow_time` / `allow_random`, resolve the range policy, precompile the
sampler; config errors raise `ValueError` at load time.
- **Range validation**: proved statically when possible. Bounds assert
by default; `clip=True` opts random draws into truncation.
- **`sample(t, rng, size)`**: random leaves sampled through sympy's
numpy registry (guarded import, symbolic fallback) around a lambdified
deterministic skeleton.
- **`evaluate(t)`**: a deterministic expression over an array of times
in one lambdified call. `sample(t=...)` is too slow to integrate a rate.
First caller: `RateSchedule` in #829.
- **`bounds`**: provable `(lower, upper)` by interval... (continued)

328 of 365 new or added lines in 1 file covered. (89.86%)

12045 of 14467 relevant lines covered (83.26%)

0.83 hits per line