Repository navigation
fix: prevent double-credit in execute_transfer for SEND txs - #1600
Conversation
|
Warning Rate limit exceeded
Your organization is not enrolled in usage-based pricing. Contact your admin to enable usage-based pricing to continue reviews beyond the rate limit, or try again in 11 minutes and 29 seconds. ⌛ How to resolve this issue?After the wait time has elapsed, a review can be triggered using the We recommend that you space out your commits to avoid hitting the rate limit. 🚦 How do rate limits work?CodeRabbit enforces hourly rate limits for each developer per organization. Our paid plans have higher rate limits than the trial, open-source and free plans. In all cases, we re-allow further reviews after a brief timeout. Please see our FAQ for further information. ✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
…it on SEND `execute_transfer` (consensus/base.py) was crediting the recipient via unconditional `update_account_balance`. For SEND transactions created by `sim_fundAccount`, the endpoint already credits the recipient via `credit_tx_value_once` (PR #1582 fix). The consensus worker then picks up the PENDING SEND and credits again in `execute_transfer` — giving 2x the funded amount. This bug has existed since Sept 2024 but was hidden by the pre-#1595 `test_accounts_funding` assertion comparing strings ("0x7d0" == 1000, which always fails inconclusively). The hex-parse fix in #1595 finally makes the test compare real values and surfaced the bug (assert 2000 == 1000). Fix: swap update_account_balance for credit_tx_value_once in execute_transfer. This is a no-op for faucet txs (value_credited already true) and a one-time credit for regular user-to-user SEND. Fixes integration test_accounts_{funding,burn,transfers} regressions.
3119d9d to
0bbd4e5
Compare
|
🎉 This PR is included in version 0.117.3 🎉 The release is available on GitHub release Your semantic-release bot 📦🚀 |
* fix(rpc): expose value_credited so execute_transfer guard works PR #1582 added immediate-credit semantics to sim_fundAccount via credit_tx_value_once. PR #1600 added an idempotency guard in execute_transfer to prevent the consensus worker from crediting again: existing_tx = transactions_processor.get_transaction_by_hash(...) if existing_tx and existing_tx.get("value_credited"): dispatch FINALIZED; return But _parse_transaction_data never included value_credited in its dict. So existing_tx.get("value_credited") always returned None, the guard never fired, and execute_transfer always credited the recipient via update_account_balance — on top of the endpoint's credit. Result: sim_fundAccount(addr, 1000) → balance 2000. Symptom in CI: integration tests test_accounts_funding, test_accounts_transfers, test_accounts_burn all asserted balance == expected, got expected*2 (or fund + 800 instead of fund - 200). Verified locally: failing tests pass with the fix; the new regression test in transactions_processor_test fails without the fix and passes with it. * ci(docker): retry apt-get update to ride out Ubuntu mirror sync hiccups Three CI runs in a row failed at the docker build stage with: E: Failed to fetch http://security.ubuntu.com/ubuntu/dists/noble-security/ main/binary-amd64/Packages.gz File has unexpected size (2017749 != 2018102). Mirror sync in progress? Ubuntu's noble-security mirror occasionally serves stale Packages.gz files during a sync rotation, persisting for several minutes. Wraps apt-get update in a 5x/15s-delay retry loop in all three backend Dockerfiles (backend, consensus-worker, database-migration) so transient mirror skew doesn't block CI. * test(integration): serialize mock-validator tests to fix flake under -n 8 Tests in tests/integration/icontracts/ that use the setup_validators fixture create mock validators with test-specific mock_response payloads. With pytest-xdist -n 8 parallelism, validators created by different tests live in the same global pool — when one test runs consensus, the pool may include validators with another test's mock data. Validators with non-matching mocks return wrong outputs, disagreement spikes, the tx errors, the test fails non-deterministically. Symptom: test_llm_erc20 (and others) fail with execution_result=ERROR and "json.loads(final_result)" stack traces; vote distributions like 4 disagree / 1 agree. Fix: - Add a pytest_collection_modifyitems hook in the icontracts conftest that tags every test using setup_validators with xdist_group(name="mock_validators"). pytest-xdist's --dist loadgroup routes tests with the same group to the same worker, serializing them. - Pass --dist loadgroup to the parallel test command in CI. Net effect: mock-validator tests run sequentially within one xdist worker; all other tests still parallelize across the remaining workers. * test(integration): fix test_llm_erc20 mock key never matching prompt The mock_response was keyed on "The balance of the sender", a string that no longer appears anywhere in the LLM prompt the contract issues. That key historically came from the eq_principle criteria, which back when the test was written was passed as part of the LLM prompt. The contract has since refactored to use `gl.eq_principle.strict_eq`, which runs leader/validators on identical inputs and verifies result equality without sending the criteria to the LLM. Result: backend/node/llm.lua's get_mock_response_from_table fell through every key without matching, then returned the literal string "no match" (its no-default fallback). Inside the contract, `json.loads("no match")` raised, GenVM exited with an error receipt, validator votes diverged depending on whose execution happened to match the leader, the tx errored, and the test failed non-deterministically. Fix: key the mock on "" — llm.lua:52-54 reserves the empty-string key as the explicit default and returns its value for any prompt. The mock now correctly returns the expected JSON for every LLM call this contract makes. * Revert "test(integration): fix test_llm_erc20 mock key never matching prompt" This reverts commit dfce5da. * test(integration): use a real prompt substring as test_llm_erc20 mock key The mock_response key "The balance of the sender" was a string that no longer appears anywhere in the LLM prompt the LlmErc20 contract issues. It used to come from gl.eq_principle's criteria back when criteria was concatenated to the LLM prompt; the contract has since moved to gl.eq_principle.strict_eq, which compares results without sending criteria to the LLM. backend/node/llm.lua matches mock_response keys against the prompt via string.find; a non-matching key falls through to the real LLM provider, whose output is non-deterministic. That's the source of the long-standing test_llm_erc20 flake (json.loads failing on either the literal "no match" sentinel or some real-LLM response). Fix: key the mock on "transactions between users" — a substring the contract's prompt opens with ("You keep track of transactions between users and their balance in coins..."). The mock now reliably returns the canned valid-JSON response for every LLM call this contract issues.
…fer moves value execute_transfer read value_credited with a plain SELECT, credited the recipient unconditionally, then set value_credited=true with no "AND value_credited = false". Any credit that landed between the read and the worker's commit (sim_fundAccount's credit_tx_value_once, as in the rc11 test_accounts_funding 2000-for-1000 failure) was applied twice. #1815 closes the endpoint side; this closes the worker side for every caller. execute_transfer now claims the credit first with the same compare-and-set credit_tx_value_once uses (claim_tx_value_credit) and only debits/credits when it won the row. A lost claim finalizes as a no-op transfer, like the old guard. Debit-then-credit ordering is unchanged; on insufficient balance the claim is released so the row stays value_credited=false and refund_tx_value semantics are unchanged. Lock order: claiming the tx row before inserting a new recipient's current_state row deadlocked against credit_tx_value_once, which inserts the account first. execute_transfer now ensures the recipient row first (ensure_account_exists, shared by both paths), then claims. Keeps #1748's atomic debit/credit and #1600's retry idempotency. Tests: tests/db-sqlalchemy/test_execute_transfer_credit_claim.py covers the endpoint credit racing in mid-transfer, the worker running against an uncommitted endpoint credit, both concurrently, the insufficient-balance release, and a retried transfer. Against the old execute_transfer three fail: two credit 2000 for a 1000 SEND, and the mid-transfer endpoint credit is not blocked by the worker.
Summary
`execute_transfer` in consensus/base.py unconditionally credited the recipient with `update_account_balance`. For SEND transactions created by `sim_fundAccount`, the endpoint already credits the recipient via `credit_tx_value_once` (PR #1582). The consensus worker then picks up the PENDING SEND and credits again — giving 2× the funded amount.
Why this was latent
The bug has existed since Sept 2024 (`c4c2d77c7`, original consensus code). It was hidden because `test_accounts_funding` asserted on strings (`"0x7d0" == 1000`) which always failed inconclusively. The hex-parse fix in #1595 finally made the test compare real values and surfaced the bug (`assert 2000 == 1000`).
PR #1582 made the endpoint idempotent but didn't touch the consensus side.
Fix
Swap `update_account_balance` for `credit_tx_value_once` in `execute_transfer`:
Tested
Follow-up (out of scope)
The sender-debit side in `execute_transfer` (line 675) is also not idempotent. For faucet txs it's skipped (`from_address=None`) but user-to-user SEND could double-debit on retry. Worth addressing separately.
Refs: #1582, #1595, #1599