Consolidate the halo2_snark slow proof tests around shared and golden proofs #3463

Activity

1. jpraynaud

   added

   to-groom 🤔Needs groomingNeeds grooming

   cryptography 🔐Cryptography relatedCryptography related

   testing 🔁Something related to testsSomething related to tests

   on Aug 4, 2026
2. jpraynaud

   added a parent issue on Aug 4, 2026

   • Implement new testing strategy for SNARK #3384
3. jpraynaud

   removed

   to-groom 🤔Needs groomingNeeds grooming

   on Aug 5, 2026
4. jpraynaud

   marked this as blocked on Aug 7, 2026

   • Reuse the cached unsafe SRS in the SNARK test setups #3433
5. hjeljeli32

   self-assigned this

   on Aug 17, 2026
6. hjeljeli32 commented on Aug 18, 2026

   

   hjeljeli32

   on Aug 18, 2026

   Last edited by hjeljeli32

   Collaborator

   More actions

   Dear @jpraynaud @damrobi

   The cached proof for different_parameters_prove_and_verify_fails does not pay

   Measured locally, ci-tests-snark, --test-threads 1:

   	that test	tier
   Before	4.624 s	15.310 s
   Cached, cold	5.273 s (+0.65)	16.085 s
   Cached, warm	4.150 s (−0.47)	14.962 s

   Proving is only ~0.74 s of that test. The other ~3.6 s is the two build_for_test calls, which the cache cannot avoid: the verifying keys still have to be derived from the current circuit to check the cached proof against. So it removes the cheapest part of the test and adds the derivations to the rest.

   On CI the cache directory is empty on every run, so this is a pure +0.65 s regression until the test caches are persisted.

   I left the test unchanged and the item unticked; the patch is ready if you disagree. Are you fine with this conclusion?
7. hjeljeli32 commented on Aug 19, 2026

   

   hjeljeli32

   on Aug 19, 2026

   Collaborator

   More actions

   Dear @jpraynaud and @damrobi, here is the gain

   Change	Scope	Before	After	Gain
   Verify negatives routed to the golden proof	4 tests, 1 deleted as a duplicate	9.8 s	0.2 s	-9.7 s (-98 %)
   Prepare-and-check and codec tests routed	4 tests	9.1 s	0.2 s	-9.0 s (-98 %)
   Wrong key read from the embedded production key instead of a second prover	1 test	3.5 s	0.05 s	included above
   Fresh-proof prepare_and_check coverage added back	1 test	2.82 s	2.84 s	+0.03 s
   Forged proof read from the content-keyed cache	1 test	4.6 s	5.3 s	not shipped, regression

   Suite	Before	After	Gain
   halo2_snark::proof slow, local warm	33.9 s	15.3 s	-18.5 s (-55 %)
   halo2_snark::proof slow, local cold	65.8 s	47.1 s	-18.7 s (-28 %)
   halo2_snark::proof slow, CI	1358 s	1140 s	-218 s (-16 %)

   Local figures are serial (--test-threads 1); the CI row compares PR #3483 against the nightly on 58b575afb, the branch's base commit.

   Remarks

   • The CI baseline for these 16 tests is 1358 s, not the 3458 s in the issue: that figure predates Reuse the cached unsafe SRS in the SNARK test setups #3433's key cache. The tier was already about 2.5x cheaper than the estimate assumed, so the saving is proportionally smaller than the ~35 minutes stated.
   • On CI the change removed 324 s outright: the seven routed tests fall from 302.7 s to 4.3 s, and the deleted one was 25.8 s. The aggregate improves by 218 s rather than 324 s because the eight untouched tests billed +106 s more than in the nightly. They are byte-identical to main, the differences are spread across them and one is negative, so this is CI wall-clock variance rather than added work.
   • The cold floor does not move: 31.8 s before and after, locally. It is the five distinct parameter sets, each paying one unsafe SRS generation plus keygen on first touch, and all five are still required by tests this ticket keeps. We removed proving, not keygen. Retiring parameter sets rather than proofs is where the remaining time sits, and that would change what the surviving proving tests prove.
8. hjeljeli32

   closed this as completedin #3483on Aug 20, 2026