2026-03-12 Engineering Team

[joecastiglione]

Agenda

• Status of skip-on-fail #1023
  • Awaiting decision from Product on default, or just merge and revise default later if needed?
  • Situation with rounding sample rates
• Status of EET merge #1042
  • Unsolved sharrow error appearing in parking location choice for SANDAG example model...
• UV documentation
• PopulationSim #217 needs attention (review)
• Testing, draft component test for the park-and-ride scenario #1001

@ActivitySim/engineering

[JoeJimFlood]

I have an update on the UV task I'd like to briefly mention (that includes a suggestion on how to improve the documentation), and I'd like to bring some attention to PopulationSim pull request #217, which is a code change needed for the implementation we set up for CDOT.

[jpn--]

ActivitySim Development Meeting Notes

Meeting Overview

Agenda

1. EET (Explicit Error Terms) feature branch merge status
2. "Skip on Fail" pull request — default setting and sample rate rounding
3. UV documentation update for installing specific ActivitySim versions
4. PopulationSim geography nesting fix
5. Draft component test review (deferred)
6. Sharrow memory leak investigation update

Participants

• Jeff Newman (@jpn--) — Driftless
• David Hensle (@dhensle) — RSG
• Sijia Wang (@i-am-sijia) — WSP
• Joe Castiglione (@joecastiglione) — Zephyr Foundation
• Jan Zill (@janzill) — Outer Loop
• Joe Flood (@JoeJimFlood) — CS
• Andrew Kay (@andkay) — CS
• Bo Wen (@bwentl) — TransLink
• Lisa Zorn (@lmz) — MTC

Duration

~65 minutes

---

Executive Summary

The meeting focused on progress and decisions around several open pull requests in ActivitySim. Jeff Newman reported near-completion of the EET feature branch merge, with a remaining bug in the parking location choice model when Sharrow is enabled; the group agreed to merge the branch anyway since the bug does not block ongoing ET work. Discussion of the "skip on fail" PR led to a concrete plan for handling divide-by-zero issues from zero sample rates and removing legacy sample rate rounding as a separate PR. Joe Flood raised two items: a UV documentation clarification for installing specific ActivitySim versions, and a PopulationSim pull request to support C-geography at a level other than second-in-hierarchy.

---

Meeting Notes

EET Feature Branch Merge

Jeff Newman reported substantial progress merging the explicit error terms (EET) branch into the main ActivitySim repository. A pull request is open to consolidate Matt's earlier work plus subsequent updates from main. One outstanding failure exists in the two-zone example model, specifically in the parking location choice component when Sharrow is enabled. The failure is deep in the model run (~90 minutes into CI, ~10 minutes locally), making it slow to debug iteratively.

Jeff proposed merging the branch despite the failing test, since: (a) it is a feature branch, not main; (b) ET testing can proceed on the legacy (non-Sharrow) code path, which is working correctly; and (c) the prior round of ET testing used a pre-Sharrow SANDAG model anyway.

Jan Zill offered to look into the failure as well. The group agreed this approach is reasonable and will not impede ongoing work.

"Skip on Fail" PR — Sample Rate Rounding and Divide-by-Zero

The "skip on fail" feature pull request has been reviewed by Jeff and David. The remaining open issue concerns sample rate rounding: historically the household sample rate was rounded to three decimal places, which caused divide-by-zero errors in some cases when weights were rounded to zero during reweighting.

Jeff proposed simply removing the rounding altogether rather than changing to a larger rounding precision. The group agreed. This will cause cascading test failures (test targets were based on rounded values), to be handled as a separate PR.

Divide-by-zero protection: Discussion explored the case where an exogenously provided sample rate of zero causes an infinite expansion factor. The agreed approach:

• Apply a numpy-style filter: if sample rate is zero, treat the inverse as zero (not infinity); if non-zero, use the normal inverse.
• Throw a warning when zero sample rates are encountered.
• Do not reweight those households.

The rounding removal will be a separate PR from the divide-by-zero fix to avoid blocking Sijia's current work.

Default setting for skip-on-fail: The product team has been asked to weigh in on whether the default should be true or false. This is a one-line change and can be resolved before a release even if the current PR is merged. The default setting does not affect existing test results either way.

Tracing functionality: David noted that Sijia's unit test should also assert that tracing values are written out for failed choices, to ensure this functionality doesn't regress. Sijia agreed to add this.

UV Documentation — Installing a Specific ActivitySim Version

Joe Flood noted that a partner agency tested their implementation using UV but had issues because their environment was configured before ActivitySim disallowed extra settings in the settings file. The workaround was to install version 1.2.1. Joe proposed adding a short note to the UV installation documentation explaining how to install a specific version of ActivitySim (e.g., pip install activitysim==1.2.1), alongside the existing instruction that installs the latest.

Jeff asked Joe to open a documentation issue/request. Joe agreed.

PopulationSim — Geography Nesting Fix

Joe Flood described a pull request he submitted for PopulationSim. The issue: the current implementation requires the C-geography (sample households) to be the second level in the geographic hierarchy (one level below meta). For CDOT's implementation, data availability required it to be at the third level, which crashed the code. Joe added a few lines in several places to support this.

Jeff noted the PR currently lacks tests and asked Joe to look into writing them. Jeff also asked whether RSG could review the PR given their PopulationSim experience. David noted that RSG's primary PopulationSim expert (Nick Fournier) recently moved to MTC. Lisa Zorn said she would ask Nick if he has time to review, and the group asked her to assign him on GitHub.

Sharrow Memory Leak Investigation

Jeff summarized ongoing work to diagnose a memory growth issue in Sharrow under multiprocessing. His hypothesis is that Numba compiled functions are immortal within a process — once loaded, memory cannot be freed. Each subprocess loads its own copy of many compiled functions, so the total footprint could scale with the number of subprocesses.

Jeff asked Bo Wen to re-run the memory profiling test with half as many threads, to determine whether total memory usage drops by roughly half (pointing to compiled function overhead) or stays similar (pointing to a data-related cause). Bo will report back next week.

Bo noted that the memory creep appeared after a prior fix to the park-and-ride MOT choice block, and speculated that garbage collection may have been triggered more frequently when memory usage was higher, masking the creep previously.

Deferred: Draft Component Test Review

This agenda item was deferred to next week, as it was not ready for discussion.

Open PRs Flagged

David flagged several other open PRs for Jeff's attention:

• Dependabot PRs (went active unexpectedly).
• Adding an ignores list to the skim load component (using Sharrow); last failing test was recently resolved.
• Disaggregate accessibility PR — code is done, tests can be added; ready for review.

---

Action Items

• Jeff Newman — Continue debugging the parking location choice failure under Sharrow + ET; push fix to the feature branch once found.
• Jeff Newman — Review open PRs: skim load ignores list and disaggregate accessibility.
• Jeff Newman — Look up Nick Fournier's current GitHub handle for the PopulationSim PR review.
• Sijia Wang — Remove the sample rate rounding (separate PR from current skip-on-fail work).
• Sijia Wang — Add divide-by-zero protection to the sample rate reweighting logic (warn and exclude zero-rate households).
• Sijia Wang — Add assertion to the skip-on-fail unit test to verify that tracing values are written out for failed choices.
• Sijia Wang / Jeff Newman — Coordinate on fixing or relaxing test targets after rounding removal causes cascading test failures.
• Joe Flood — Open a GitHub issue/PR to add UV documentation for installing a specific version of ActivitySim.
• Joe Flood — Investigate writing tests for the PopulationSim C-geography nesting PR.
• Lisa Zorn — Ask Nick Fournier if he can review the PopulationSim C-geography PR; assign him on GitHub.
• Bo Wen — Re-run Sharrow memory profiling with half the number of threads and report results next week.
• Product team — Weigh in on the default setting (true or false) for skip-on-fail before next release.