Conversation
This was referenced Sep 15, 2026
Contributor
|
Unisay
force-pushed
the
yura/policy-filters-costing
branch
from
September 15, 2026 14:06
0cc03d0 to
3f85fb2
Compare
Unisay
force-pushed
the
yura/policy-filters-costing
branch
from
September 15, 2026 14:17
3f85fb2 to
b1883c1
Compare
Unisay
force-pushed
the
yura/policy-filters-costing
branch
from
September 15, 2026 15:54
b1883c1 to
9655366
Compare
Unisay
force-pushed
the
yura/policy-filters-costing
branch
from
September 15, 2026 16:02
9655366 to
ac15cf8
Compare
Unisay
force-pushed
the
yura/policy-filters-costing
branch
from
September 16, 2026 10:45
ac15cf8 to
731b32f
Compare
Unisay
force-pushed
the
yura/policy-filters-costing
branch
from
September 16, 2026 15:46
731b32f to
de2f3ea
Compare
Unisay
force-pushed
the
yura/policy-filters-costing
branch
from
September 17, 2026 13:27
de2f3ea to
d2de75c
Compare
`Value` gains a fifth cache, the number of negative amounts per currency, so that removing a whole currency subtracts a cached count instead of recounting its amounts. That recount was the only cache-maintenance term proportional to the number of amounts rather than to the number of currencies, and the one a cost model cannot see. The cache is maintained by `insertCoin`, `unionValue`, `scaleValue` and `unValueData`. Their memory models account for the extra field, and their CPU models are refitted from benchmarks that hold negative amounts throughout: the rows they were fitted on never held one, so they never paid for the extra search path, and on inputs that do the previous `insertCoin` model undercharged every row, by up to 1.55x, and `scaleValue` three rows in four, by up to 1.46x. Benchmarked on the reference machine.
Looking each id up in the outer map and inserting what it finds costs p log m whatever the list is. Building a `Set` of the list first cost p log p however small the `Value`, and folding the restricted map through `pack'` recounted the negative amounts of every currency kept; with the per-currency counts the caches now come from the kept currencies alone, so both filters are logarithmic in the `Value`.
`ValueOuterDepth` measures the depth of the outer map and nothing else. Neither filter descends an inner map, so `ValueMaxDepth` would over-measure a `Value` of one policy and thousands of tokens by almost the whole measure. The empty `Value` measures 1 rather than 0, so that a product-shaped model cannot charge a policy list its intercept alone.
One generator for both builtins, since both walk the policy list once at one outer-map descent per element. A random sample over the plane of list length and outer-map depth, plus sweeps that isolate the depth term, check that the tokens per policy do not show, and vary what the model cannot see. Each family is defined next to its own description; see Note [Benchmarking keepPolicies and dropPolicies] for what they have in common.
CPU and memory are both `multiplied_sizes` in the length of the policy list and the depth of the `Value`'s outer map, eight new parameters. The CPU models take `fit.fan`'s intercept and the smallest slope that covers every benchmark row rather than nine in ten: the grid is the whole domain the builtins accept, and the cost of a descent keeps growing with the depth up to its edge, so the rows `fit.fan` would leave out are the deepest maps a script can present. Benchmarked on the reference machine: https://github.com/IntersectMBO/plutus/actions/runs/34979629812
Two interactive pages under doc/cost-models plotting the benchmark rows and the shipped model on the plane of list length and outer-map depth, so the charge can be read against the measurement point by point. The pages read the model from the cost-model JSON rather than refitting it. The shared script gains what reading a two-argument model needs: the model drawn as a translucent surface over the whole plane (the default) or as a cross at each benchmark point, linear or logarithmic axes, a table of how many points the model over- and undercharges with the median and worst factor on each side, and coefficients that can be edited in place so a candidate fit can be tried by hand and read off the same table and surface. The two existing three-dimensional pages, `unionValue` and `valueContains`, use the same controls. A page can be pointed at explicit CSV and JSON files with `?csv=...&json=...`, e.g. a checkout served locally.
Both filters now fail on a `Value` with more than `policyFilterMaxSize` policies. Their models charge on the length of the list times the depth of the outer map, and past that depth a descent into an outer map that no longer fits in cache outgrows the shape; the bound makes those inputs unreachable by construction rather than by an argument about protocol parameters. The benchmark grid stops at the bound, since a row beyond it would fail instead of timing a descent.
Unisay
force-pushed
the
yura/policy-filters-costing
branch
from
September 18, 2026 09:57
d2de75c to
39934e8
Compare
Unisay
marked this pull request as ready for review
September 18, 2026 09:57
Contributor
|
📊 Cost model pages for this PR: https://IntersectMBO.github.io/plutus/pr-preview/cost-models/pr-7954/?branch=yura%2Fpolicy-filters-costing |
The pages read the benchmark CSV and the cost model JSON from a branch of this repository and remember the last branch used in the browser, so a preview link without a branch can plot another branch's data. The preview comment now carries `?branch=<head ref>`, that parameter wins over what the browser remembers, and links between pages keep it.
Unisay
force-pushed
the
yura/policy-filters-costing
branch
from
September 18, 2026 13:31
272a2ee to
3034334
Compare
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
1 participant
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.
Adds the cost models for
keepPoliciesanddropPolicies: CPU and memory are bothmultiplied_sizesin the length of the policy list and the depth of theValue's outer map, eight new parameters. To make that shape sound,Valuegains a fifth cache, the count of negative amounts per currency, which removes the only cache-maintenance term proportional to the number of amounts rather than the number of currencies; both filters become O(p log m).The cache is maintained by
insertCoin,unionValue,scaleValueandunValueData, whose benchmarks never held a negative amount and so never paid for it; they now do throughout, and the CPU models are refitted from a run on the reference machine.keepPoliciesmultiplied_sizes {355913, 46004}multiplied_sizes {32, 32}dropPoliciesmultiplied_sizes {356925, 57953}multiplied_sizes {32, 36}Both filters fail on a
Valuewith more than 8191 policies (policyFilterMaxSize), which is three to four times the most policies a 16 KB transaction can carry;assetCount v <= 8191is a free conservative guard for callers. The pages underdoc/cost-modelsdraw a two-argument model as a translucent surface, tabulate how many points it over- and undercharges and by how much, and let the coefficients be edited in place to try a candidate fit against the same table.Collapses #7930, #7937 and #7946 into one PR. Implements IntersectMBO/plutus-private#2310 and IntersectMBO/plutus-private#2311.