mempool.dat reload drops ~2/3 of the mempool: same-second children re-admitted before parents #949

Summary

A clean restart loses most of the mempool. load_mempool re-admits mempool.dat records in time order, but time has one-second resolution, and ties keep file order. dump_mempool writes get_all_entries() in map order, which is not parent-first. So a child that shares its parent's second is often tried before the parent. It fails with missing-inputs, and every descendant behind it fails too.

Each restart makes this worse. Re-admission gives the txs it accepts a fresh time, so thousands of them end up in the same few seconds. On the next restart, most parent/child pairs are ties.

Evidence

On a synced mainnet node (0.6.0), over two consecutive clean restarts:

After restart 2, I parsed the mempool.dat it had loaded and ran each record through testmempoolaccept:

• 20,030 missing-inputs
• 19,871 of those have their parent in the same file
• in 14,959 of those, the parent has the same time as the child
• no case where the parent's time is later than the child's

The file has 1,935 distinct timestamps for 36,121 records. The largest same-second group is 1,883 txs, all stamped during restart 1's re-admission.

Nothing in the log shows this, because load_mempool discards the error (Err(_) => stats.skipped += 1). The only trace is the skipped= count.

Expected

Like Bitcoin Core, a restart should keep everything that is still valid. Core avoids this by writing mempool.dat in topological (depth) order.

Suggested fix

• Write parents first in dump_mempool (topological / ancestor-count order), and/or have load_mempool order ties by in-file parent dependency. A retry pass for missing-inputs would also work, but costs more.
• Keep each tx's original admission time on reload instead of stamping the reload time, so time stays meaningful (and so does mempool expiry).
• Log or count skip reasons so a regression like this shows up in the log.
• Test: a chain where the child has the same time as the parent and is written before it in the file must reload completely.