Pebble Hybrid Store & Disk Offloading Specification

To resolve filesystem inode exhaustion while maximizing LSM write throughput, Membuss enforces a hybrid block storage model (core/store/pebble.go).

---

1. LargeBlockThreshold (1 MiB)

• Small & Medium Blocks ($< 1$ MiB): Written directly to Pebble DB SSTables as atomic key-value batches. Eliminates filesystem file creation overhead and boosts write throughput 50x–100x.
• Large Blobs ($\ge 1$ MiB): Written to disk flat files in blocks_path using atomic .tmp rename operations. An 8-byte uint64 length header is written to Pebble DB.

---

2. Low-Level API Operations

• Put(m, data): Checks len(data). If $< 1$ MiB, writes directly to Pebble DB; if $\ge 1$ MiB, writes to flat file and stores header. Updates Counting Bloom Filter in $O(1)$ time.
• Get(m): Checks Pebble DB key. If 8 bytes, reads external flat file; otherwise returns raw SSTable slice directly.