Posts about building a hybrid database engine - the parts that worked, the parts that didn't, and what we learned shipping them.
MongrelDB ships full-text relevance as an ordinary SQL function that scores whole Arrow batches per call, composes with WHERE and ORDER BY and window functions, and asks for no index up front, with the honest tradeoff that global IDF lives in the fts_docs virtual table instead.
MongrelDB stored procedures are versioned, checksummed definitions installed over HTTP, and the pure PHP client exposes the whole lifecycle as five typed calls whose name-keyed argument arrays map directly onto the procedure's declared params.
MongrelDB Kit stamps every applied migration with a content-aware SHA-256 checksum over its recorded intent and re-verifies it on every run, so editing a historical migration after it shipped fails loudly at the next deploy instead of silently re-shaping production.
The MongrelDB PHP client mirrors the daemon's two auth modes with one constructor argument each, and anything more exotic, from per-request token vending to signing proxies, goes behind a one-method TransportInterface you control.
MongrelDB Kit's typed schema DSL and its raw SQL surface are two front doors to the same engine rather than one compiling into the other, and knowing which door to use, and how to mix both inside a single migration, is most of the skill.
MongrelDB Kit ships TypeScript, Rust, and Python SDKs over a single Rust core, and a shared JSON conformance suite proves all three produce identical results and identical error codes on top of one shared query compiler.
The MongrelDB PHP client substitutes malformed UTF-8 instead of rejecting the whole request, and reserves hard errors for values like INF, NAN, and recursive structures that genuinely have no JSON representation.
The MongrelDB PHP client treats schema as typed data instead of SQL strings, so the same call that creates a table can be paired with a catalog read that proves the server stored what you think it stored.
The MongrelDB PHP client's fluent query builder names the index family in each condition instead of generating SQL for a planner to second-guess, so the index choice you made at schema time is the one that executes.
The MongrelDB PHP client stages multi-row writes locally, commits them as one atomic HTTP batch, and dedupes retries with storage-layer idempotency keys, so a timed-out request is safe to replay without building an outbox.
The MongrelDB PHP client maps HTTP status codes to a small typed exception hierarchy, so a 401 is an AuthException, a 409 is a ConstraintException carrying the server's error code and the failing operation index, and your catch blocks read like the failure modes they handle.
The MongrelDB PHP client hides all of HTTP behind a one-method TransportInterface, ships cURL as the default with a stream fallback, and lets you inject a fake for tests or a PSR-18 bridge for Guzzle without touching application code.
The MongrelDB PHP client opts into PHP 8.5's persistent cURL share handles to carry DNS, TLS sessions, and the connection pool across FPM requests, and silently falls back to per-request pooling on 8.4.
A walk through the actual HTTP surface mongreldb-server exposes, from /sql and /txn to sessions, pagination cursors, and the /kit/* typed routes, and why plain HTTP was the right transport bet.
MongrelDB ships as one compiled binary per platform and one shared library per ABI, with cluster and cloud integrations excluded at compile time, and that shape is an economic decision as much as an engineering one.
MongrelDB evaluates unique, foreign-key, and CHECK constraints atomically at transaction commit instead of per statement, which changes what a batch import can promise and what it has to fear.
A technical comparison of MongrelDB and Couchbase Lite for embedded documents, offline sync, edge vector search, transactions, analytics, and security.
A technical comparison of MongrelDB and MongoDB Atlas for documents, transactions, vector and full-text search, hybrid retrieval, deployment, and encryption.
A technical comparison of MongrelDB and Qdrant for dense and sparse vectors, filtering, hybrid search, transactions, deployment, and local RAG.
A technical comparison of MongrelDB with PostgreSQL, pgvector, and ParadeDB for transactions, hybrid search, analytics, deployment, and operations.
A technical comparison of MongrelDB and DuckDB for embedded analytics, transactions, vector and full-text search, Arrow workflows, and mixed workloads.
A technical comparison of MongrelDB with Turso and libSQL for embedded replicas, SQLite compatibility, vector search, analytics, synchronization, and security.
A technical comparison of MongrelDB and ObjectBox for embedded applications, on-device vector search, transactions, synchronization, analytics, and encryption.
A technical comparison of MongrelDB and LanceDB for embedded vector search, hybrid retrieval, columnar data, transactions, RAG, and local deployment.
A technical comparison of MongrelDB and SurrealDB across embedded deployment, data models, transactions, SQL, vector search, hybrid retrieval, and operations.
A technical comparison of MongrelDB and SQLite for transactions, analytics, vector and text search, encryption, deployment, and local-first applications.
An end-to-end technical walkthrough of MongrelDB, from transaction validation, WAL durability, MVCC, and PAX columnar runs to DataFusion SQL and hybrid retrieval.
A JSON column in MongrelDB is a typed column that the engine validates at write time and queries from the inside, so the schema-versus-schemaless argument stops being a binary you have to pick a side of.
A column type is not documentation, it is a promise the storage engine can spend, and MongrelDB's type system covers fixed-width integers, decimal128 money, five flavors of time, sortable UUIDs, native JSON, and typed arrays so the engine can plan bytes instead of guessing at strings.
A pure embedded engine never has to think about other processes, but MongrelDB is also a server, so it takes an exclusive advisory lock at open, fails fast by default, waits politely when you ask it to, and pushes genuine multi-process access through the daemon instead of through the filesystem.
Most database auth lives in middleware, which means every code path that skips the middleware skips the auth. MongrelDB stores Argon2id-hashed users and roles in the catalog and can enforce permissions on every transaction at the storage layer itself.
MongrelDB reconstructs change events from its own WAL and streams committed records to follower copies, so an embedded database can participate in replication and change data capture without bolting on a separate toolchain.
MongrelDB binds idempotency keys into the commit log itself, so a retried write is safe across process crashes and restarts, not just across network drops.
MongrelDB separates accepted writes from fsync-backed durable commits, batches WAL work with group commit, and flushes log-structured columnar runs later.
PHP 8.4 is the floor of the pure PHP client, and PHP 8.5, released in November 2025, is where the language gives us enough to stop writing workarounds for cURL and readonly cloning.
MongrelDB runs in-process or behind mongreldb-server, keeping one storage and query engine while changing ownership, authentication, serialization, and resource boundaries.
SQLite SEE is the official paid reference for SQLite encryption. Here is MongrelDB's open-source AES-GCM design, key hierarchy, WAL coverage, and search tradeoff.
MongrelDB's PGM learned-range index uses an epsilon-bounded piecewise linear model to predict sorted-key positions, then verifies the bounded range exactly.
Install the MongrelDB PHP 8.4+ client with Composer and connect to mongreldb-server over HTTP without compiling a custom PHP extension.
How an embedded vector database keeps HNSW candidates, operational rows, SQL filters, sparse retrieval, and exact reranking inside one Rust engine.
MongrelDB uses a Burrows-Wheeler transform and wavelet-tree FM-index for exact substring candidates that combine with SQL, vectors, ranges, and equality filters.
MongrelDB supports 35 languages through nine embedded native bindings and 26 language-native HTTP clients, with one engine contract under both boundaries.
An honest map of where MongrelDB fits among embedded databases: operational writes, columnar scans, hybrid retrieval, encryption, SQL, and single-node limits.
