How BLite implements transaction isolation using two in-memory dictionaries — a WAL cache for uncommitted writes and a WAL index for committed-but-not-checkpointed pages — and where that falls short of true MVCC.
> Read moreHow BLite encodes the file type, collection slot, and local page number into a single uint — and why this self-describing page ID lets the storage engine route reads and writes across multiple memory-mapped files with a bitmask and no schema lookup.
> Read moreHow BLite implements the HNSW algorithm for nearest-neighbor vector search — level assignment, greedy layer traversal, Algorithm 4 neighbor selection, and the .NET specifics that keep it allocation-light.
> Read moreA walkthrough of BLite's B+Tree implementation: page-aligned nodes, struct-based headers, node splitting with ArrayPool, range scans over doubly-linked leaf pages, and the stackalloc trick in the read path.
> Read moreHow BLite uses a Write-Ahead Log to guarantee durability, and the Group Commit pattern — backed by System.Threading.Channels and TaskCompletionSource — that amortizes the cost of fsync across concurrent transactions.
> Read moreHow BLite translates LINQ expressions into B-Tree page lookups — ExpressionVisitor, predicate merging, hybrid in-memory fallback, and the async streaming trick you can't do without Task.Run.
> Read moreTwo engineering challenges in BLite: supporting private fields, private constructors and nested collections without runtime reflection — and managing serialization buffers without putting pressure on the GC.
> Read moreWhy I replaced BSON's null-terminated field names with 2-byte integer IDs, how the dictionary system works, and an honest critique of the trade-offs I made.
> Read moreAn introduction to MrDevRobot: my personal space for architecture deep-dives, open source stories, and technical writing.
> Read more