Nowcast can't archive a new round: the hosted forecast_uk is pinned to boe-var data ending 2026Q1

未关闭
  • 哪里不对？
#224 3 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 9 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
34/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
github-actions, python
领域
backend, devops

调研方向

Read nowcast.py and the workflow that invokes it, then trace how it calls hosted forecast_uk and archives rounds. Check whether a cheap capabilities or metadata request can detect the hosted model's data edge before the expensive forecast call; the issue also reports a redirect failure in mcp/client/streamable_http.py. Done means avoiding forecast calls when no newer round is available and checking the endpoint on the first due run after #195 merges.

由索引模型根据 Issue 内容生成。

描述

data question

Problem

Even once the data store is refreshed (#195 via #221), the nowcast job cannot archive a new round. The round comes from forecast_uk on the hosted MCP server, and that model does not read macro's data store. It reads boe-var-model's bundled data/boe_var_data.csv at the commit pinned in pyproject.toml:

  • boe_var @ git+https://github.com/PolicyEngine/boe-var-model@42cbcbb…
  • at 42cbcbb (and on boe-var-model main, last data commit cf5fdf7 on 2026-07-20, "Refresh data vintage to 2026Q1 (#2)"), the last row of data/boe_var_data.csv is 2026Q1

So nowcast.py will stop with data edge 2026Q1 is already archived and nothing new has been released; no round written, as it did on 13 Aug (run 31670728758) and 19 Aug. The record has had no new round since 2026-07-29. The two archived boe-svar rounds, and the 07-28 unemployment satellite round, all have data edge 2026Q1.

There is also an unrelated crash: on 16 Sep (run 35058603658) the call itself failed with httpx.HTTPStatusError from mcp/client/streamable_http.py::_raise_for_unfollowed_redirect. That points to the Modal URL https://policyengine--policyengine-macro-mcp-serve.modal.run/mcp answering with a redirect. I have not re-tested it, to avoid starting a Modal container.

Upcoming releases a round could be archived ahead of

GDP 2026Q3 first estimate 12 Nov, unemployment 17 Nov, CPI 18 Nov (the next_release dates in #195's data).

Cost if nothing changes

Once #195 merges, --due 3 is true on 9–12 Nov and again on 14–18 Nov. On each of those days the job installs the MCP client and calls forecast_uk with 6000 draws, which takes about 5.5 minutes on a cold Modal container, only to be told "data edge 2026Q1 already archived". The same thing happened on 12–13 Aug and 17–19 Aug.

Options

  1. Refresh boe-var-model's data to 2026Q2 before 12 Nov: run scripts/download_data.py, re-pin macro's boe_var, and redeploy the MCP app. A round with a 2026Q2 edge that forecasts 2026Q3 is a scoreable claim for the 12/17/18 Nov releases only if it is archived before 07:00 on 12 Nov. That is a model-data change plus a deploy, so it needs you. A refresh after 12 Nov gives a 2026Q3 edge, which cannot be scored against these releases; its first scored period would be Q4 (February).
  2. Skip the November window. The existing rounds already cover 2026Q3–2029Q2 and will be scored as the outturns land. Stop the wasted Modal calls by also gating nowcast.py on the hosted model's data edge: a cheap capabilities/metadata call, or comparing against the last archived edge before calling forecast_uk. After that, make the boe-var data refresh a standing task right after each GDP first estimate, so the next round (edge 2026Q3, archived before the February releases) is in place.
  3. Either way: after #195 merges, look at the first due run to check the endpoint no longer redirects (16 Sep crash).

Recommendation: option 2 now, without the gate the job will spend roughly 9 cold Modal calls in November for nothing. Choose option 1 only if you can refresh, re-pin and deploy calmly by 11 Nov.

主要语言
HTML
星标
3
派生
0
平均合并
8 天 15 小时
30 天内合并 PR
1

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

PolicyEngine/macro 的其他 Issue

查看 PolicyEngine/macro 的全部 Issue

相似的 Issue

更多 Backend & API Design Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。