iotaledger / iota / 28987808076

69%

Build Type

push

github

Committed by web-flow

Commit Message

fix(core): use on-chain timestamps for smooth pruning (#12132)

# Description of change

Full nodes accumulated far more on-disk state than expected while
syncing (observed on testnet: ~840 GB perpetual store, ~400 GB
checkpoints store) even with pruning correctly configured.

Root cause: pruning was **smoothed against wall-clock time**.
`smoothed_max_eligible_checkpoint_number` advanced the prune watermark
by only `delta = (max_eligible - pruned) / num_intervals` per tick,
where `num_intervals = epoch_duration_ms / pruning_tick_duration_ms`
(e.g. `86_400_000 / 10_000 = 8640` for a 24h epoch). At equilibrium this
leaves an unpruned backlog of roughly `num_intervals ×
(eligible-growth-per-tick)`. At tip that is ~1 epoch (the intended
behavior — it spreads each epoch's delete burst to avoid I/O spikes),
but during catch-up sync execution races thousands of checkpoints ahead
per tick, so the backlog — and disk usage — balloons. The smoothing had
no notion of how far behind tip the node was.

This PR replaces wall-clock smoothing with a **chain-time sliding
window**. The prune loop already walks checkpoints forward from the
highest pruned checkpoint, loading each summary, so we simply stop when
a checkpoint is newer than a retention cutoff:

```
cutoff_ts = highest_executed_checkpoint.timestamp_ms - num_epochs_to_retain * epoch_duration_ms
prune while checkpoint.timestamp_ms <= cutoff_ts
```

Checkpoint `timestamp_ms` is consensus-agreed and monotonic, so the
cutoff lands on the same checkpoint whether history is replayed in an
hour or lived at tip — the retained span is bounded to one retention
window in all cases, and the deletes still slide smoothly (no per-epoch
burst) at tip. The change touches a single shared function, so both
object and checkpoint pruning benefit.

Notes:
- The existing epoch-count guard (`current_epoch < ckpt.epoch() +
num_epochs_to_retain`) is untouched and remains the hard correctness
bound (the state accumulator and epoch-scope... (continued)

Coverage Stats

115 of 124 new or added lines in 2 files covered. (92.74%)

96 existing lines in 21 files now uncovered.

161181 of 236688 relevant lines covered (68.1%)

463448.5 hits per line