Repository navigation
benchmarking/evaluate: run evaluations, with a run_results source #3551
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Merged
Max Smythe (maxsmythe)
merged 3 commits into
agent-substrate:main
from
roycaihw:run-evaluations-b
Nov 12, 2026
Merged
Changes from all commits
Commits
Show all changes
3 commits
Select commit
Hold shift + click to select a range
File filter
Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,141 @@ | ||
| # Run evaluations | ||
|
|
||
| `evaluate` judges a finished benchmark run against CEL conditions declared in | ||
| its test entry, and returns the verdicts as a `Report`. | ||
| Depends on [`cel-expr-python`](https://github.com/cel-expr/cel-python) and | ||
| [pydantic](https://docs.pydantic.dev/) (`requirements.txt`, Python 3.11 or later). | ||
|
|
||
| ## Model | ||
|
|
||
| - A **source** reads the run and returns several named **values**, or | ||
| says it can't judge the run (`invalid`, with a reason). Sources: | ||
| `run_results`. | ||
| - A **variable** is one source query: `{name, source, source_cfg}`. Its values | ||
| are exposed to CEL as a map under `<variable>` (`<variable>.<value>`, or | ||
| `<variable>["<value>"]` when `<value>` contains dots or CEL keywords). | ||
| - A **condition** is `{name, cel}`: a [CEL](https://cel.dev/) boolean | ||
| expression over the variables. | ||
| - An **evaluation** is a named set of variables and conditions, AND-ed: it | ||
| `fail`s if any condition failed, else is `invalid` if any condition or | ||
| variable is invalid, else `pass`es. Evaluations are judged independently, with no hierarchy between | ||
| them. | ||
| - The run's **summary** combines its evaluations the same way. | ||
| - `passed` is true for `pass` only, and false for both `fail` and `invalid`, | ||
| so an alert should key on `passed`. | ||
|
|
||
| ## Config | ||
|
|
||
| Evaluations live in the test entry, next to the test they judge. This one | ||
| checks two envelope health gates: | ||
|
|
||
| ```yaml | ||
| - name: sweperf_150u_c3hc192metal_microvm | ||
| # ... | ||
| evaluations: | ||
| - name: envelope_health | ||
| description: "Failure ratio and worker memory within the scalability envelope" | ||
| variables: | ||
| - name: trial | ||
| source: run_results | ||
| source_cfg: {metric: trial_summary, measurements: [aggregate_failure_ratio, allocatable_ram_gb]} | ||
| - name: server | ||
| source: run_results | ||
| source_cfg: {metric: server_summary, measurements: [working_set_node_max_gb]} | ||
| conditions: | ||
| - {name: failure_rate, cel: "trial.aggregate_failure_ratio <= 0.001"} | ||
| - {name: worker_memory, cel: "server.working_set_node_max_gb < 0.85 * trial.allocatable_ram_gb"} | ||
| ``` | ||
|
|
||
| - Evaluation, variable and condition names match `^[a-z][a-z0-9_]*$`, at most | ||
| 64 characters, and are unique where they live. CEL itself rejects a | ||
| variable named after a CEL type (`int`, `map`) or a keyword (`in`, `true`) | ||
| when a condition compiles. Value names come from the source | ||
| and may be any non-empty string | ||
| (`locust_ResumeActor`, `metric_nighthawk.builtin_attempted_rps`, `p99.9`). | ||
| Unknown or non-string mapping keys are errors, and `source_cfg` holds JSON | ||
| values only, so a YAML date in it is an error too. There is no version field | ||
| while the schema is experimental. | ||
| - Conditions are standard CEL, compiled when the evaluations load, with | ||
| `round(x, digits)` added. Each must evaluate to `bool`: a bare value such as | ||
| `trial.ok` compiles, and a result that isn't a bool makes the condition `invalid`. | ||
| - Values enter CEL as doubles, whatever number type the source returns, | ||
| since they come from JSON, which has one number type, and as ints `a / b` | ||
| would be integer division. The results keep the source's own value types. | ||
| CEL has no arithmetic between int and double, so write `2.0 * t.x`, not | ||
| `2 * t.x`; the same goes for comparing a double expression | ||
| (`round(t.x, 1) >= 90.0`). A value compared directly may use an int | ||
| literal (`t.x <= 1500`). Mixing int and double fails when the evaluations | ||
| load if CEL can see both types; otherwise (`t.x * 100 >= 99`, where | ||
| `t.x` is only known at run time) the condition is `invalid` on every run, | ||
| so it never passes. | ||
| - Missing or null values are omitted from `<variable>`'s map: reading one | ||
| directly fails lookup and makes the condition `invalid`, while | ||
| `has(<variable>.<key>)` returns `false` without error. A value CEL can't | ||
| hold makes the conditions `invalid`. An `invalid` variable is left out of | ||
| the CEL data, so a condition that reads it is `invalid` with CEL's error, | ||
| and the evaluation is `invalid` even if no condition reads it: a source | ||
| failure never lets an evaluation pass. | ||
| - CEL `==` and `!=` across distinct dynamic types (`1.1 == "1.10"`, | ||
| `true == 1.0`) evaluate to `false` and `true` without raising an error. Float | ||
| division inside a condition follows IEEE-754 (`1.0 / 0.0` produces `+Inf` | ||
| and `0.0 / 0.0` produces `NaN`). | ||
|
|
||
| ## Sources | ||
|
|
||
| A source is a class with three hooks (`sources/base.py`): | ||
|
|
||
| | hook | runs | does | | ||
| |---|---|---| | ||
| | `validate(cfg)`, a classmethod | when the evaluations are loaded | raises `ValueError` for a bad `source_cfg` | | ||
| | `init(run)` | once, when the run starts (its files may not exist yet) | takes the `BenchmarkRunContext`, the run being evaluated: which run it is (`identity`: `uri`, and `test_name`, `tag`, `run_date`, `run_ts` when known), where it is stored, and its stored artifacts (`artifact(name)`, a local path or None); the source chooses what it reads | | ||
| | `query(cfg)` | after the run, once per distinct `source_cfg` | returns named values (numbers, bools, strings or null), or raises `SourceInvalid(reason, values)` when it can't judge the run | | ||
|
|
||
| Value names come from `source_cfg` or are fixed by the source, never from the | ||
| run's data. The evaluator creates one instance of each source the evaluations | ||
| name. The values of a `SourceInvalid` are kept in the results, while the | ||
| variable and the conditions reading it become `invalid`. Any other exception | ||
| from a hook makes the source's variables `invalid` with the error, and so does | ||
| a value of another type, such as a list or a mapping, or a value name that | ||
| isn't a non-empty string. Sources never write files. A new source is a class | ||
| in `sources/` and a line in `SOURCES`. | ||
|
|
||
| ### `run_results` | ||
|
|
||
| The measurements of one `stats.jsonl` row: | ||
|
|
||
| ```yaml | ||
| source_cfg: | ||
| metric: trial_summary # if repeated, the last row wins | ||
| measurements: [aggregate_failure_ratio] # returned under the same names | ||
| ``` | ||
|
|
||
| Numeric strings without `_` become floats and `"true"`/`"false"` bools. Strings | ||
| with `_` (`"2026_11"`) stay strings. A measurement the row lacks, or a number | ||
| that isn't finite (NaN, an infinity, `1e999`), is null. Blank lines are skipped; | ||
| any other line that isn't a JSON object with a `metric` makes the variable | ||
| `invalid`, naming the line, since the runner writes one such object per line. | ||
|
|
||
| ## Results | ||
|
|
||
| `Evaluator.evaluate` returns a `Report` (`evaluator.py`): the run, the | ||
| `summary` (`status`, `passed`, `failed`, `invalid`), and per evaluation its | ||
| status and reason, each condition with its CEL expression and verdict, each | ||
| variable with its values (and why it is invalid), and the evaluation as loaded | ||
| (`config`). An evaluation's reason names the conditions that failed, else the | ||
| invalid conditions and variables. | ||
|
|
||
| ## Layout | ||
|
|
||
| | file | | | ||
| |---|---| | ||
| | `config.py` | loads and validates an `evaluations` block (pydantic models) | | ||
| | `cel_env.py` | shared CEL environment, `round()`, and evaluation | | ||
| | `evaluator.py` | runs the source hooks and judges a run into a `Report` | | ||
| | `sources/` | the hooks (`base.py`) and the sources | | ||
|
|
||
| ## Tests | ||
|
|
||
| ```sh | ||
| hack/verify/benchmarking-python.sh # as CI runs them; or, with requirements.txt installed: | ||
| cd benchmarking/evaluate && python3 -m unittest discover -s unit_tests | ||
| ``` |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,21 @@ | ||
| # Copyright 2026 The Agent Substrate Authors | ||
| # | ||
| # Licensed under the Apache License, Version 2.0 (the "License"); | ||
| # you may not use this file except in compliance with the License. | ||
| # You may obtain a copy of the License at | ||
| # | ||
| # http://www.apache.org/licenses/LICENSE-2.0 | ||
| # | ||
| # Unless required by applicable law or agreed to in writing, software | ||
| # distributed under the License is distributed on an "AS IS" BASIS, | ||
| # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| # See the License for the specific language governing permissions and | ||
| # limitations under the License. | ||
|
|
||
| """Run evaluations: judge a finished benchmark run against conditions declared in its test entry. | ||
|
|
||
| A test entry's `evaluations:` (config.py) name sources, which read the run's | ||
| files into values (sources/), and CEL conditions over those values | ||
| (cel_env.py). evaluator.py judges them into a Report. Depends on | ||
| cel-expr-python and pydantic (requirements.txt). See README.md. | ||
| """ |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,77 @@ | ||
| # Copyright 2026 The Agent Substrate Authors | ||
| # | ||
| # Licensed under the Apache License, Version 2.0 (the "License"); | ||
| # you may not use this file except in compliance with the License. | ||
| # You may obtain a copy of the License at | ||
| # | ||
| # http://www.apache.org/licenses/LICENSE-2.0 | ||
| # | ||
| # Unless required by applicable law or agreed to in writing, software | ||
| # distributed under the License is distributed on an "AS IS" BASIS, | ||
| # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| # See the License for the specific language governing permissions and | ||
| # limitations under the License. | ||
|
|
||
| """Shared CEL environment, round(), and evaluation.""" | ||
|
|
||
| from typing import Any, Mapping | ||
|
|
||
| from cel_expr_python import cel | ||
|
|
||
|
|
||
| def _round(x: float, digits: float) -> float: | ||
| if not float(digits).is_integer(): | ||
| raise ValueError(f"round: digits {digits} isn't a whole number") | ||
| return round(x, int(digits)) | ||
|
|
||
|
|
||
| # round(x, digits) as Python rounds, so an expression can keep a fixed number of decimals. | ||
| ROUND = cel.FunctionDecl("round", [ | ||
| cel.Overload(f"round_double_{t.name().lower()}", return_type=cel.Type.DOUBLE, parameters=[cel.Type.DOUBLE, t], | ||
| impl=_round) for t in (cel.Type.DOUBLE, cel.Type.INT)]) | ||
|
|
||
|
|
||
| def to_cel(values: Mapping[str, Any]) -> dict[str, Any]: | ||
| """values as CEL reads them: ints as doubles, and no key for None, so has() | ||
| tests presence. Source values come from JSON, which has one number type, and | ||
| as ints `a / b` would be integer division.""" | ||
| return {k: float(v) if isinstance(v, int) and not isinstance(v, bool) else v | ||
| for k, v in values.items() if v is not None} | ||
|
|
||
|
|
||
| def evaluate_condition(expr: cel.Expression, data: Mapping[str, Any]) -> bool: | ||
| """Whether the boolean expr is true for data. Raises ValueError with CEL's | ||
| error when it fails to evaluate, for example on a key data lacks or a value | ||
| CEL can't hold.""" | ||
| try: | ||
| result = expr.eval(data=data) | ||
| except RuntimeError as e: # a value CEL can't take in, such as a string with a lone surrogate | ||
| raise ValueError(str(e)) from None | ||
| if result.type() == cel.Type.ERROR: | ||
| raise ValueError(str(result.value())) | ||
| if result.type() != cel.Type.BOOL: | ||
| raise ValueError(f"gives a {result.type().name().lower()}, not a bool") | ||
| return result.value() | ||
|
|
||
|
|
||
| class Env: | ||
| """A CEL environment with variables and functions.""" | ||
|
|
||
| def __init__(self, variables: Mapping[str, cel.Type], functions: tuple[cel.FunctionDecl, ...] = ()): | ||
| self.variables = dict(variables) | ||
| try: | ||
| self.env = cel.NewEnv(variables=self.variables, functions=list(functions)) | ||
| except RuntimeError as e: | ||
| raise ValueError(str(e)) from None | ||
|
|
||
| def compile(self, expression: str, where: str, | ||
| gives: tuple[cel.Type, ...] = (cel.Type.BOOL,)) -> cel.Expression: | ||
| """expression compiled. Raises ValueError unless it type-checks and gives one of gives.""" | ||
| try: | ||
| expr = self.env.compile(expression) | ||
| except RuntimeError as e: # a syntax or type error, or a variable CEL already declares | ||
| raise ValueError(f"{where}: {e}") from None | ||
| if expr.return_type() not in gives: | ||
| wanted = " or a ".join(t.name().lower() for t in gives if t != cel.Type.DYN) | ||
| raise ValueError(f"{where}: must give a {wanted}, gives {expr.return_type().name()}") | ||
| return expr | ||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,124 @@ | ||
| # Copyright 2026 The Agent Substrate Authors | ||
| # | ||
| # Licensed under the Apache License, Version 2.0 (the "License"); | ||
| # you may not use this file except in compliance with the License. | ||
| # You may obtain a copy of the License at | ||
| # | ||
| # http://www.apache.org/licenses/LICENSE-2.0 | ||
| # | ||
| # Unless required by applicable law or agreed to in writing, software | ||
| # distributed under the License is distributed on an "AS IS" BASIS, | ||
| # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| # See the License for the specific language governing permissions and | ||
| # limitations under the License. | ||
|
|
||
| """The `evaluations:` block of a test entry: loading and validation. | ||
|
|
||
| evaluations: | ||
| - name: envelope_health # unique in the test | ||
| description: "..." # optional | ||
| variables: # each is one source query, with several named values | ||
| - {name: trial, source: run_results, source_cfg: {...}} | ||
| conditions: # AND-ed | ||
| - {name: failure_ratio, cel: "trial.aggregate_failure_ratio <= 0.001"} | ||
|
|
||
| The models below give the shape. Loading also checks what a shape can't: that | ||
| names are unique, each variable's source_cfg (by its source's validate hook), | ||
| and that each condition compiles, so a bad config fails before a run is read. | ||
| There is no version field while the schema is experimental. | ||
| """ | ||
|
|
||
| from typing import Annotated, Any | ||
|
|
||
| from cel_expr_python import cel | ||
| from pydantic import Field, JsonValue, PrivateAttr, TypeAdapter, ValidationError, field_validator, model_validator | ||
|
|
||
| from . import cel_env | ||
| from .sources import SOURCES | ||
| from .sources.base import Model, Name, describe | ||
|
|
||
| CelExpression = cel.Expression | ||
|
|
||
|
|
||
| class ConfigError(ValueError): | ||
| """The evaluations block is invalid; the message names where.""" | ||
|
|
||
|
|
||
| class Variable(Model): | ||
| name: Name | ||
| source: str | ||
| # JSON, as the results echo it: a YAML date or a non-string key is an error here. | ||
| # Validation builds new containers, so later edits to the input don't reach it. | ||
| source_cfg: dict[str, JsonValue] = {} | ||
|
|
||
| @field_validator("source") | ||
| @classmethod | ||
| def _known(cls, source: str) -> str: | ||
| if source not in SOURCES: | ||
| raise ValueError(f"source {source!r} is not one of {sorted(SOURCES)}") | ||
| return source | ||
|
|
||
| @model_validator(mode="after") | ||
| def _source_cfg(self) -> "Variable": | ||
| try: | ||
| SOURCES[self.source].validate(self.source_cfg) | ||
| except ValueError as e: | ||
| raise ValueError(f"source_cfg: {e}") from None | ||
| return self | ||
|
|
||
|
|
||
| class Condition(Model): | ||
| name: Name | ||
| cel: str | ||
| _expr: CelExpression = PrivateAttr() # compiled by the Evaluation, which knows the variables | ||
|
|
||
| @property | ||
| def expr(self) -> CelExpression: | ||
| return self._expr | ||
|
|
||
|
|
||
| class Evaluation(Model): | ||
| name: Name | ||
| description: str = "" | ||
| # Lists, not tuples: pydantic counts a tuple's length after dropping the items that failed, | ||
| # so one bad item would also report the tuple as empty. | ||
| variables: list[Variable] = Field(min_length=1) | ||
| conditions: list[Condition] = Field(min_length=1) | ||
|
|
||
| @model_validator(mode="after") | ||
| def _compile(self) -> "Evaluation": | ||
| _unique([v.name for v in self.variables], "variable") | ||
| _unique([c.name for c in self.conditions], "condition") | ||
| env = cel_env.Env({v.name: cel.Type.Map(cel.Type.STRING, cel.Type.DYN) for v in self.variables}, | ||
| functions=(cel_env.ROUND,)) | ||
| for c in self.conditions: | ||
| # DYN too, as when a condition is just a value (`trial.ok`); a non-bool | ||
| # result then makes the condition invalid when it runs. | ||
| c._expr = env.compile(c.cel, f"condition {c.name!r}: cel", # pylint: disable=protected-access | ||
| gives=(cel.Type.BOOL, cel.Type.DYN)) | ||
| return self | ||
|
|
||
|
|
||
| _EVALUATIONS = TypeAdapter(Annotated[list[Evaluation], Field(min_length=1)]) | ||
|
|
||
|
|
||
| def load(evaluations: Any) -> list[Evaluation]: | ||
| """The evaluations of a test entry, validated; raises ConfigError.""" | ||
| try: | ||
| out = _EVALUATIONS.validate_python(evaluations) | ||
| _unique([e.name for e in out], "evaluation") | ||
| except ValidationError as e: | ||
| raise ConfigError(f"evaluations: {describe(e)}") from None | ||
| except ValueError as e: | ||
| raise ConfigError(f"evaluations: {e}") from None | ||
| return out | ||
|
|
||
|
|
||
| def to_json(evaluations: list[Evaluation]) -> list[dict[str, Any]]: | ||
| """The evaluations as loaded, defaults filled: what the results echo, and load() reads back.""" | ||
| return [e.model_dump(mode="json") for e in evaluations] | ||
|
|
||
|
|
||
| def _unique(names: list[str], what: str) -> None: | ||
| if dupes := sorted({n for n in names if names.count(n) > 1}): | ||
| raise ValueError(f"{what} names repeated: {dupes}") |
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Uh oh!
There was an error while loading. Please reload this page.