Expose the query ID and the last provisional value to the cycle recovery function#1012
Merged
MichaReiser merged 3 commits intosalsa-rs:masterfrom Oct 23, 2025
Merged
Conversation
✅ Deploy Preview for salsa-rs canceled.
|
MichaReiser
requested review from
carljm and
ibraheemdev
and removed request for
carljm
CodSpeed Performance ReportMerging #1012 will not alter performanceComparing Summary
|
MichaReiser
force-pushed
the
expose-id-and-last-prov-value
branch
from
ae90661 to
9fc59a2
Compare
ibraheemdev
approved these changes
Oct 22, 2025
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
MichaReiser
force-pushed
the
expose-id-and-last-prov-value
branch
from
9fc59a2 to
620ced5
Compare
Merged
mtshiba
added a commit
to mtshiba/salsa
that referenced
this pull request
Oct 29, 2025
mtshiba
added a commit
to mtshiba/salsa
that referenced
this pull request
Oct 29, 2025
mtshiba
added a commit
to mtshiba/salsa
that referenced
this pull request
Oct 29, 2025
mtshiba
added a commit
to mtshiba/salsa
that referenced
this pull request
Oct 29, 2025
carljm
added a commit
to astral-sh/ruff
that referenced
this pull request
Nov 26, 2025
## Summary Derived from #17371 Fixes astral-sh/ty#256 Fixes astral-sh/ty#1415 Fixes astral-sh/ty#1433 Fixes astral-sh/ty#1524 Properly handles any kind of recursive inference and prevents panics. --- Let me explain techniques for converging fixed-point iterations during recursive type inference. There are two types of type inference that naively don't converge (causing salsa to panic): divergent type inference and oscillating type inference. ### Divergent type inference Divergent type inference occurs when eagerly expanding a recursive type. A typical example is this: ```python class C: def f(self, other: "C"): self.x = (other.x, 1) reveal_type(C().x) # revealed: Unknown | tuple[Unknown | tuple[Unknown | tuple[..., Literal[1]], Literal[1]], Literal[1]] ``` To solve this problem, we have already introduced `Divergent` types (#20312). `Divergent` types are treated as a kind of dynamic type [^1]. ```python Unknown | tuple[Unknown | tuple[Unknown | tuple[..., Literal[1]], Literal[1]], Literal[1]] => Unknown | tuple[Divergent, Literal[1]] ``` When a query function that returns a type enters a cycle, it sets `Divergent` as the cycle initial value (instead of `Never`). Then, in the cycle recovery function, it reduces the nesting of types containing `Divergent` to converge. ```python 0th: Divergent 1st: Unknown | tuple[Divergent, Literal[1]] 2nd: Unknown | tuple[Unknown | tuple[Divergent, Literal[1]], Literal[1]] => Unknown | tuple[Divergent, Literal[1]] ``` Each cycle recovery function for each query should operate only on the `Divergent` type originating from that query. For this reason, while `Divergent` appears the same as `Any` to the user, it internally carries some information: the location where the cycle occurred. Previously, we roughly identified this by having the scope where the cycle occurred, but with the update to salsa, functions that create cycle initial values can now receive a `salsa::Id` (salsa-rs/salsa#1012). This is an opaque ID that uniquely identifies the cycle head (the query that is the starting point for the fixed-point iteration). `Divergent` now has this `salsa::Id`. ### Oscillating type inference Now, another thing to consider is oscillating type inference. Oscillating type inference arises from the fact that monotonicity is broken. Monotonicity here means that for a query function, if it enters a cycle, the calculation must start from a "bottom value" and progress towards the final result with each cycle. Monotonicity breaks down in type systems that have features like overloading and overriding. ```python class Base: def flip(self) -> "Sub": return Sub() class Sub(Base): def flip(self) -> "Base": return Base() class C: def __init__(self, x: Sub): self.x = x def replace_with(self, other: "C"): self.x = other.x.flip() reveal_type(C(Sub()).x) ``` Naive fixed-point iteration results in `Divergent -> Sub -> Base -> Sub -> ...`, which oscillates forever without diverging or converging. To address this, the salsa API has been modified so that the cycle recovery function receives the value of the previous cycle (salsa-rs/salsa#1012). The cycle recovery function returns the union type of the current cycle and the previous cycle. In the above example, the result type for each cycle is `Divergent -> Sub -> Base (= Sub | Base) -> Base`, which converges. The final result of oscillating type inference does not contain `Divergent` because `Divergent` that appears in a union type can be removed, as is clear from the expansion. This simplification is performed at the same time as nesting reduction. ``` T | Divergent = T | (T | (T | ...)) = T ``` [^1]: In theory, it may be possible to strictly treat types containing `Divergent` types as recursive types, but we probably shouldn't go that deep yet. (AFAIK, there are no PEPs that specify how to handle implicitly recursive types that aren't named by type aliases) ## Performance analysis A happy side effect of this PR is that we've observed widespread performance improvements! This is likely due to the removal of the `ITERATIONS_BEFORE_FALLBACK` and max-specialization depth trick (astral-sh/ty#1433, astral-sh/ty#1415), which means we reach a fixed point much sooner. ## Ecosystem analysis The changes look good overall. You may notice changes in the converged values for recursive types, this is because the way recursive types are normalized has been changed. Previously, types containing `Divergent` types were normalized by replacing them with the `Divergent` type itself, but in this PR, types with a nesting level of 2 or more that contain `Divergent` types are normalized by replacing them with a type with a nesting level of 1. This means that information about the non-divergent parts of recursive types is no longer lost. ```python # previous tuple[tuple[Divergent, int], int] => Divergent # now tuple[tuple[Divergent, int], int] => tuple[Divergent, int] ``` The false positive error introduced in this PR occurs in class definitions with self-referential base classes, such as the one below. ```python from typing_extensions import Generic, TypeVar T = TypeVar("T") U = TypeVar("U") class Base2(Generic[T, U]): ... # TODO: no error # error: [unsupported-base] "Unsupported class base with type `<class 'Base2[Sub2, U@Sub2]'> | <class 'Base2[Sub2[Unknown], U@Sub2]'>`" class Sub2(Base2["Sub2", U]): ... ``` This is due to the lack of support for unions of MROs, or because cyclic legacy generic types are not inferred as generic types early in the query cycle. ## Test Plan All samples listed in astral-sh/ty#256 are tested and passed without any panic! ## Acknowledgments Thanks to @MichaReiser for working on bug fixes and improvements to salsa for this PR. @carljm also contributed early on to the discussion of the query convergence mechanism proposed in this PR. --------- Co-authored-by: Carl Meyer <carl@astral.sh>
Merged
This was referenced Dec 17, 2025
Merged
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
Extend the
cycle_recoveryfunction with two new arguments:id: The query id. Can be used as a cheap unique identifier to "re-identify" a value returned by a previous iterationlast_provisional_value: The value returned in the last iteration. Useful to detect if the value is divergingI included two more changes in this PR. Happy to split them out in their own PR if desired:
cycle_fnoptional and provide a default that returnsIteratebecause this is enough for most cyclic functionscycle_fnreturnsFallback, compare the new value with the last provisional. If they're the same, consider this query as converged (there's no point in running another iteration).