Question about long-running Binance Spot local order books

I bumped into a problem whereas testing long-running Binance Spot local order e-book caches and needed to match notes with individuals who build towards the Binance API.

Benchmark setup:

- BTCUSDT

- 25.10 hours

- similar WebSocket stream

- naive implementation vs actively pruned implementation

The naive version still seemed effective at greatest bid / greatest ask, however the full native e-book degraded over time:

- 24.09% bid match

- 39.82% ask match

- 21,244 orphaned native levels

The pruned model ended at:

- 87.83% bid match

- 91.74% ask match

- 305 orphaned local levels

This isn't meant as “Binance docs are flawed”. The present Binance guide paperwork limit=5000 snapshots and update-ID gap handling.

My question for individuals sustaining their very own Binance Spot local order books:

Do you implement a retention boundary / pruning policy, or do you rely only on qty=0 updates to take away levels?