Repository navigation
2026-09-10 Engineering Team #138
Description
Agenda
- Phase 11 closeout and release readiness: confirm release scope, remaining blockers, and timing; coordinate version/beta status with Product/Community.
ActivitySim
- Final review / merge decisions:
- #1103 Faster RNG: Jan approved today; address minor comments and confirm performance/memory validation.
- #1112 Park-and-ride documentation: assign review; include with the upcoming release?
- Low-memory strategy:
- #1110 Reduce peak memory: assign independent review and full-scale testing.
- #1106 Automatic adaptive chunking: initial technical assessment and recommendation for Product/Community.
- Resolve owner blockers:
- #1107 Calibration: estimation-test failures, user documentation, and review readiness.
- #1083 Shadow-pricing segmentation: Sijia's unit tests and SANDAG validation.
- #1062 Summarize memory cleanup: David/RSG's test and completion status.
- Guide closeout: #1086 Application and Analysis Guide: resolve review-access issues, confirm SANDAG walkthrough testing, and set final feedback deadline.
- Skim loading: #1108 OMX/Zarr precedence: agree intended behavior and assign ownership.
- Proposal: ZARR is a convenience cache only and is not to be modified; if both Zarr and OMX inputs exist they are assumed to contain identical content. Zarr is not an agreed industry data transfer standard and lacks widespread vendor support.
- Recently merged: 🥳
PopulationSim
- #219 Stabilization: remaining review/testing items and merge decision; unblock #217 Intermediate geographies.
- #221 Household-weight caps: Amir/Nick testing results and next fixes.
Performance-testing infrastructure
- #1078 Performance-test subarea: Joe Flood's runner/data PR, warning cleanup, and agreed runtime/memory benchmarks.
Coordination
- Modeling Mobility: decide joint submission and technical topics; assign authors ahead of the September 19 deadline.
Activity
jpn-- commentedon Sep 18, 2026
Meeting Overview
Agenda
Participants
Duration
Approximately 62 minutes.
Executive Summary
The group agreed to proceed with the reviewed RNG implementation after clarifying its error reporting, without adopting an additional AI-generated rewrite that maintainers could not confidently understand. The upcoming ActivitySim release will include the RNG changes and park-and-ride documentation, with other pending work deferred to later releases. The group confirmed that user-facing calibration documentation already exists in the calibration PR. Memory improvements in PR #1110 reportedly allow the CTPS model to run with approximately 52–53 GB of peak memory, but require careful review of chunking behavior and compatibility across models.
Meeting Notes
Faster RNG and AI-generated core code
Jeff proposed replacing the names “simple,” “fast,” and “faster” with names that identify the implementations:
legacy,PCG64, andSFC64 hash. The existing names could imply a performance ordering that does not always hold.Jan’s review identified a concern about reading a fixed-width block from a NumPy RNG state address. In exploring alternatives, Jeff used AI to develop a rewrite that removed reliance on NumPy internals and CFFI. An initial version introduced an approximately 10% runtime regression; a subsequent version reportedly recovered performance and produced bit-for-bit identical results, but introduced low-level Numba/LLVM and architecture-specific code that Jeff could not confidently maintain.
The discussion emphasized that:
Jan also questioned whether retaining EET sampling justified the added complexity and suggested considering a simpler set of supported options. Jeff explained that repeatedly using NumPy’s standard seeding approach would impose substantial runtime costs under ActivitySim’s reproducibility requirements. No decision was made to remove EET sampling.
Decision: Jeff will clarify the memory-layout error message and merge the reviewed RNG implementation without the additional low-level AI-generated rewrite.
Park-and-ride documentation and release scope
Park-and-ride functionality is already on the main branch, but its documentation was omitted. Jeff prepared documentation for the model components and lot choice functionality, reviewed it briefly, and requested David’s review for accuracy and usability.
David committed to reviewing the documentation that day.
Decision: Merge the RNG changes and reviewed park-and-ride documentation, then issue a release within days. This will allow agencies, including TfNSW, to share an installable released version with stakeholders.
Other pending changes, including memory improvements, shadow-price segmentation, and calibration work, will be considered for later releases. No additional release-blocking engineering requirements were raised.
Calibration documentation
Jan relayed an earlier request for searchable, user-facing Sphinx documentation, an example setup, and concise guidance alongside the YAML configurations users edit.
David showed that the calibration PR already includes user-guide documentation. It is not yet visible in the published documentation because the PR has not been merged and built.
The group agreed that this addressed the documentation concern. The calibration PR itself remains unfinished and is outside the immediate release scope.
Memory reductions in PR #1110
Jeff described PR #1110, developed to help the CTPS Boston model run on a computer with 64 GB of RAM.
The main change moves preprocessing inside explicit chunk loops, particularly for destination-choice logsums. Previously, preprocessing could create temporary columns for the entire chooser table before chunking began, limiting the memory savings from chunking. In some components, joins between preprocessing and utility evaluation also needed to move inside the chunk loop.
The PR also reduces table copies and attempts to release unused process memory across operating systems.
Jeff reported that:
Review requirements
David and Jan questioned whether existing CI tests exercise the relevant chunking paths. They also raised concerns about preprocessing expressions that depend on information outside a chunk, such as household-level trip counts.
Reviewers need to determine whether chunk boundaries preserve the necessary household context and whether calculations remain valid when preprocessing operates on subsets of the data.
Jeff requested review from Jan, David, CJ, and Andrew, including consideration of client models outside CI coverage, such as MWCOG and ARC models. Jan agreed to continue his review.
David requested before-and-after memory profiles and timing logs to help assess the changes. Jeff agreed to locate and share them.
Further optimization opportunities
The PR addresses the components needed to meet the CTPS memory target; it does not optimize every ActivitySim component.
Jeff identified repeated Sharrow flow selection and hashing within chunk loops as a possible source of avoidable overhead. Reusing the appropriate compiled flow across chunks could improve runtime, but this optimization has not been implemented or fully measured.
Review direction: Correctness concerns should be resolved before merging. Additional optimization opportunities should be recorded as follow-up issues rather than blocking the PR solely because more improvements are possible.
Deferred topics
Action Items