Reuse Execution for repeated CalculationHandle calls

[viktar-b]

Problem

A Python module can call estimate through the generated handle. Each call makes a new Execution object. That procedure uses 8.6 ms. The arithmetic uses 0.001 ms.

A test outside the repository measured 1000 calls of the two-panel estimate.

• Plain Python, 1000 calls. Total 0.0007 s. Mean 0.001 ms per call.
• CSO handle, 1000 calls. Total 8.63 s. Mean 8.63 ms per call.
• Execution.__init__. Mean 7.48 ms.
• Planner.plan. Mean 3.64 ms.
• run() on the same engine after observations.clear(). Mean 0.16 ms.

The same engine gave the correct results for width=2 and then width=1.

Cause

CalculationHandle.__call__ makes a new Execution for each call from ordinary Python. Execution.__init__ reads the source, makes a plan, changes the AST, and compiles the code. The generated binding already keeps one handle for the process.

estimate = load_calculation("../estimate.cso.py", function="estimate", fingerprint="...")

The code is in packages/cso-python/src/cso_python/handles.py.

Recommended change

Keep the Execution object on the handle after the first call.

For the next calls, clear observations and call run() with the new inputs.

Make a new Execution if the set of input names is different from the plan. Make a new Execution if the captured source files changed.

Do not change the public call.

from _cso_bindings.estimate import estimate

result = estimate(width=2)

Do not remove tracing. The default call must continue to record observations.

This change does not decrease the time of python -m cso_python execute. That command starts a new process for each run.

Acceptance

• 1000 calls of the two-panel estimate through the generated handle complete in less than 1 s.
• Results stay the same for width=2 and width=1.
• If the source files change, the next call uses the new source.
• If the input names change, the next call makes a new plan.