Skip to content

feat(chronos_forecasting): 1.0.0 on the shared forecasting interface - #196

Merged
Aliaksei-Kharlap merged 1 commit into
mainfrom
feat/chronos-forecasting-1.0
Oct 30, 2026
Merged

Aliaksei-Kharlap merged 1 commit into
mainfrom
feat/chronos-forecasting-1.0

Conversation

@Aliaksei-Kharlap

@Aliaksei-Kharlap Aliaksei-Kharlap commented Oct 30, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

chronos_forecasting 1.0.0: the reference implementation of the shared forecasting interface for InfluxDB 3 Explorer.
The request, response, errors and stored rows change, so this is a major release.

  • Forecast a measurement field, or the result of any SQL query, from any database (source_database)
  • tag_filter, sql_filter, group_by; every series goes to one model call
  • series_limit with series_total in every response; results come in a stable order
  • History: window, point_limit, sampling_interval, a bounded default window; history laid on a regular grid; a series mixing tag sets that write at the same time is refused
  • Horizon as horizon_steps or horizon_duration, capped by max_forecast_points; levels give lower_<percent> / upper_<percent>
  • On-demand backtest (backtest_points): coverage per level and MAE, in the same model call
  • Settings from env, trigger arguments, TOML, request body, headers and query string (influxdata-plugin-utils); unknown keys are refused
  • Error codes with HTTP statuses (400 / 500 / 503); too_few_points per series
  • Stored rows carry points_read, run_time and an optional forecast_id tag
  • Shared model cache with load locking and LRU, safe under concurrent runs; the model loads only when a series can be forecast
  • forecast_interface.json manifest for Explorer; new dependency influxdata-plugin-utils>=0.5.0
  • README split by trigger type; TOML templates for the scheduled and HTTP triggers; metadata with http_body_config

Breaking changes from 0.1.x

  • Renamed keys: table/measurement → source_measurement, where_clause → sql_filter, tag_values → tag_filter, context_limit → point_limit, agg_interval →
    sampling_interval, horizon → horizon_steps, model_id → model
  • The response is results[] per series × field; point key median → forecast
  • A horizon is required; unknown keys are refused; errors answer with HTTP 4xx/5xx
  • Stored rows: the full model id, a source_measurement tag, no mode tag

Testing

  • Live test plan of 227 cases on InfluxDB 3.12.1 Enterprise: all pass
  • influxdb3-plugin package (SDK 0.6.0): OK; the artifact includes `forecast_interface.json

Closes #184
Closes #185 (items 1–5; the CPU build of PyTorch cannot be chosen through install_packages)
Refs #186, influxdata/influxdb3_ui#1817

@Aliaksei-Kharlap
Aliaksei-Kharlap merged commit 6eac959 into main Oct 30, 2026
7 checks passed
@Aliaksei-Kharlap
Aliaksei-Kharlap deleted the feat/chronos-forecasting-1.0 branch October 30, 2026 11:47
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

chronos_forecasting: batching, matching inputs and smaller installs for Explorer chronos_forecasting: forecast a query, in any database

2 participants