dowel

Architecture

This is a design document about internals. As a user, the behavior you can rely on is covered by 60-cli.md (commands and the output contract) and 63-guides.md (working with the cache).

1. One core, multiple frontends

Incremental evaluation, typed values with provenance, and the language server are not separate features but three aspects of the same core — the same structure as rust-analyzer sitting on the same query foundation (Salsa) as rustc.

• The incremental evaluation engine = what the language server needs (partial re-evaluation per keystroke)
• Typed values with provenance = what the language server needs (go-to-definition, hover, diagnostics)
• dowel build and the language server pose different queries to the same query graph

So “build the build system first, add the language server later” is not a viable order. The constraints that cannot be retrofitted are fixed first; shipping the language server itself is deferred.

2. The four constraints that cannot be retrofitted

Write the evaluator as a naive tree-walking interpreter and the following become unrecoverable:

1. Error-tolerant parsing — do not stop at syntax errors; keep partial trees and continue evaluating. The lossless CST is the source of truth; the AST is a projection of it
2. Spans everywhere — every value carries a source location, surviving string expansion and transformation
3. Cancellability — every keystroke cancels the previous query; every layer of the evaluator must propagate cancellation
4. Guaranteed termination — the reason for non-Turing-completeness is not simplicity of authoring but guaranteeing the language server always responds

3. Query catalog (initial draft)

Query	Input	Output
parse(file)	file contents	CST
eval(module)	CST, parent scope	value environment
resolve(target)	name	target definition
interface(target)	target	merged propagated properties
probe(toolchain, check)	toolchain hash, check contents	fact
plan(target, config)	target, configuration	action graph

Required optimizations

• Early cutoff — if re-evaluation produces the same result as before, do not invalidate dependents. This is why a comment edit does not cascade into action-graph regeneration; it is the highest-value optimization in Salsa-style implementations
• Durability layering — manifests change often; toolchain facts almost never. Layers carry durability levels, and re-validation of stable layers is skipped entirely
• Parallel evaluation and cancellation

4. Value representation

Value = { type, data, provenance }

Provenance is a constituent of the value, not side-band data. Since provenance is a projection of the query graph, it costs almost nothing extra once the incremental engine exists.

5. Persistence (no daemon)

There is no resident daemon (ADR-0002). Since the in-memory graph cannot be kept across processes, three things substitute:

5.1 Make restoration cost O(touched nodes)

Avoid serialize-everything / restore-everything — it forfeits the incremental advantage as the graph grows.

• An mmap-able fixed-length record index + an append-only value log
• The index holds only query-key hashes, output fingerprints, and offsets of dependency edges
• Node bodies are not read until needed. Validation is fingerprint comparison only, so most nodes are judged “unchanged” without reading their values
• mmap plus the OS page cache is the effective substitute for residency; most of the cost of avoiding a daemon is recovered right here

5.2 Change detection

File watching is unavailable, so changes are judged by a stat sweep over the known input set.

• The key is (mtime, size, inode, ctime); content is hashed only when they differ
• At a few thousand files, parallel stat takes single-digit milliseconds — the territory ninja has already proven
• mtime granularity and clock skew on network file systems are handled by falling back to hashing
• Validation propagates lazily: leaf changes are only marked, and dependents are validated when a query actually reaches them

5.3 Concurrent access from multiple processes

Required because the CLI and the language server touch the same store.

• The value log is append-only; in-progress writes are invisible to everyone
• Index updates are written to a temporary file and swapped in with an atomic rename
• The writer is limited to one (flock). A process that cannot take the lock computes in its own memory and writes nothing back. Correctness is never lost — only the cached speedup
• Invariant: a process dying at any point never corrupts the store

5.4 What this drags in

• A startup-time budget — target: under 10ms when there is nothing to do. Without a daemon, startup is paid on every run, so implementation-language choice, avoiding dynamic linking, and limiting files read at startup all matter
• Store GC — append-only means growth. The store lives under .dowel/cache/ (gitignored) and dowel cache gc collects by generation count or size cap
• A separate probe-fact DB — toolchain-dependent facts should be shared across projects, so they live in the user cache area, content-addressed; the top of the durability hierarchy

6. Where the language server stands

“No daemon” and a language server do not conflict. The distinction is who starts it and how long it lives.

	Started by	Lifetime	User’s perception
daemon	implicit	outlives projects	unaware it exists; unclear how to stop it
language server	the editor	ends with the editor	something explicitly started

Invariants:

• The CLI never depends on the language server’s existence. Every feature produces the same result whether or not it is running
• The language server’s in-memory graph is derived from the disk store; the disk is always the source of truth
• The language server does not hold the writer lock persistently, and results derived from unsaved buffers are never written to the store

7. Intended scale and limits

A daemonless design loses at extreme scale. Bazel / Buck2 adopt daemons because at hundreds of thousands of nodes, index validation itself dominates.

The intended scale is 10^3–10^4 targets. Whole-monorepo single-graph usage is not a goal.

8. Relation to a future execution layer

Content addressing, append-only logs, and fingerprint validation are exactly the machinery an action cache needs; extending to one later reuses them as-is. Nothing gets built twice.

9. Reducing cold configure

Configure time decomposes into three terms:

1. Probe execution — try_compile and friends; each one launches a compiler and linker process
2. Discovery sweeps — file-system walking by find_package / pkg-config
3. Manifest evaluation and file writing

The felt slowness comes mostly from 1 and 2, independent of implementation language. 3 dominates only at thousands of targets.

The countermeasure: treat probe results not as an implicit cache (CMakeCache.txt-style) but as an independent fact database.

• Keyed by toolchain hash + probe source + flags
• Shareable across projects and machines; as long as the same compiler is used, the same probe runs once
• Cross compilation becomes swapping the fact DB, shrinking hand-written cross files
• It also matters for reproducibility: today’s probes depend on “the state of the host they ran on”, an unrecorded input, which this promotes to an explicit one

This site is open source. Improve this page.