The chaindb_wipe test suite runs after chaindb_backend_selection and
rocksdb_wrapper, both of which leave the process-wide static g_rocksdb
(and on some paths the leveldb txdb singleton) alive. A leaked
g_rocksdb means the next test that does MakeChainDB('cr+') may get a
path that the prior test's open handle is still serving — leading to
the test operating on stale state and the on-disk wipe having no
effect. The H1 crashed_migration_marker_triggers_retry test
specifically could not bootstrap a fresh txleveldb/ for the migration
because the leveldb handle from the prior test was still bound.
This is the same class of bug as W2 (live LevelDB iterator outliving
the DB close) but at the test binary's process-lifetime scale: a live
DB handle from a prior test leaks into the next test and the on-disk
wipe is a no-op.
Fix: add a ResetChainDBStatics() helper that explicitly opens + closes
both backends (in create-if-missing mode so it works whether or not a
prior test left a DB on disk) and then wipes the on-disk chain DB
directories. Call it at the top of every chaindb_wipe test.
Before: 2/4 chaindb_wipe tests passing (H1 retry, H4 happy path) due
to the static-state leak. The crashes were also producing spurious
SIGABRTs at process exit from the static VersionSet assertion.
After: 20/20 chaindb_runtime tests pass, 3/3 chaindb_equivalence,
14/14 snapshotnet.
One file, +52 lines, no production code changes.
Three fixes for the chain-DB migration path on real chain data.
All three were uncovered when running the full DNS2 2.2M-block chain
end-to-end; the existing 18 unit tests passed because they exercised
small fixtures, never the real migration entry point.
W2 (root cause): chaindb_migrate.cpp — scope the source.NewIterator()
inside an inner block so it's destroyed BEFORE source.Close(). Live
LevelDB iterators hold a Version ref; closing the DB with one alive
trips the dummy_versions_.next_ == &dummy_versions_ assertion in
leveldb::VersionSet::~VersionSet (version_set.cc:755), aborting the
daemon after verification but before the marker is removed. This
explains the original H4 symptom: the daemon died in the gap between
'verified' and 'fs::remove', and Release builds hid it by compiling
asserts out. The H1 retry path's static-state issue in the test
binary is the same bug at process exit. In-loop failures now break
out with fCopyOK=false and are handled after the iterator dies.
H4 (defense in depth): chaindb_migrate.cpp — keep the verify-and-fail
hardening even though W2 fixes the cause. Use the non-throwing
error_code overload, fs::exists verify after remove, single 100ms
retry (Windows AV/indexer transient locks), hard-fail strError if
the marker still survives. Operator-visible failure beats silent
re-migration time bomb. The H4 invariant: a successful migration
never leaves the marker on disk.
W1: init.cpp — Lookup('0.0.0.0', addrBind, GetListenPort(), false)
replaced with direct CService construction from in_addr{htonl(INADDR_ANY)}.
This was the bug that prevented fc7ad5b from ever starting on
SAMI-PC; Windows getaddrinfo doesn't always map the literal '0.0.0.0'
string to INADDR_ANY.
Test: chaindb_runtime_tests.cpp — adds marker_removed_after_successful_migration
which exercises the real MaybeMigrateLevelDbToRocksDb() end-to-end on
the happy path. Complements the existing
crashed_migration_marker_triggers_retry (retry path). This is the
gap that hid the original bug: no test went through the production
entry point on the happy path.
Runtime verification: full DNS2 chain state (txleveldb 1.1GB +
blk0001.dat 942MB, 6.77M records) migrated end-to-end. MIGRATION_INCOMPLETE
absent from disk after. Reopened rocksdb reads back cleanly via
getblockcount / LoadBlockIndex.
Three files, 152 insertions, 27 deletions, build clean, CI ready.
Adds a Boost test that pre-creates a rocksdb/ dir with a
MIGRATION_INCOMPLETE marker and verifies the init path:
1. Detects the marker
2. Refuses to open the rocksdb/ dir as live state
3. Re-runs the migration
Also exercises the M4 marker flush+verify path: marker is on
disk only during an in-progress migration and removed on success.
This test was in the H1/H2/H3/M4 patch but never committed;
folding it in here so the test surface matches the audit doc.
H1: init.cpp now detects crashed migrations (MIGRATION_INCOMPLETE marker)
and retries instead of opening a partial RocksDB. Refuses to start if
the marker persists after migration attempt.
H2: LevelDB ExistsRaw now returns false for keys deleted in the active
batch, matching ReadRaw and the RocksDB backend. Fixes latent
cross-backend consensus split in intra-batch spend checks.
H3: All RocksDB close paths now go through close_rocksdb() which
destroys CF handles before deleting the DB. Fixes RocksDB assertion
/ UB on shutdown and version-reset.
M4: Migration marker write is now flushed + verified (refuses to start
migration if marker can't be written).
CF routing permanently disabled: GetCF() always returns nullptr (default
column family). The read path (NewIterator, LoadBlockIndex) only iterates
the default CF, so writes routed to per-prefix CFs were invisible to scans.
This is why -chaindb=rocksdb compiled clean but was never runtime-valid.
Existing CF-enabled DBs still open (handles retained for cleanup) but no
routing occurs. CF-aware iteration is a future follow-up.
txdb-factory: RocksDB is now the default backend (was still leveldb).
Tests updated for RocksDB-as-default expectations.
From Claude's ROCKSDB-T010-REVIEW-2026-07-01 audit on E:\repos\triangles.
- txdb-factory.cpp: drop static-cache in ResolveChainDbKind so the
-chaindb flag can be toggled at runtime (needed for tests; cost is
negligible since the daemon sets it once at startup)
- txdb-rocksdb.cpp: fix ExistsRaw to honor pending-batch delete markers.
Previously a key erased inside an open batch was still reported as
existing because the underlying DB hadn't been updated yet. Mirror
ReadRaw's correct behavior: a delete marker shadows the DB value.
- chaindb_runtime_tests.cpp: per-test fresh handle via close-reopen
dance so the static g_rocksdb singleton doesn't leak state between
cases. Tests filter framework keys (length-prefixed 'version' and
'dbformat') from iterator walks. block_index test fixed to Seek()
not Seek("blockindex") since the serialized keys start with the
length byte 0x0a.
- snapshotnet_tests.cpp, chaindb_runtime_tests.cpp: include wallet.h,
ui_interface.h, uint256.h, checkpoints.h as needed for linker; add
BOOST_TEST_MODULE decl; define global stubs (pwalletMain,
uiInterface, fConfChange, etc.) so wallet.cpp link succeeds.
Result: test_snapshotnet + test_chaindb_runtime both pass with zero
errors. Found and fixed a real production bug in ExistsRaw along
the way.
- snapshotnet.cpp: always re-scan on HasServableSnapshot; auto-dump
from current chain when synced to canonical snapshot height
- net.cpp: EnsureLocalSnapshot() at startup so NODE_SNAPSHOT reaches
outbound peers in the first version message
- CMakeLists.txt: add test_snapshotnet + test_chaindb_runtime targets
- test/snapshotnet_tests.cpp, test/chaindb_runtime_tests.cpp: full
coverage for the SnapshotNet P2P protocol + CRocksTxDB wrapper layer
Krystie