RFC: SQLite-based state management for ToolHive #41
Conversation
Proposes replacing file-based local state management with a single SQLite database. Phase 1 targets skills to unblock skills lifecycle management. Later phases will migrate workload statuses and groups. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Juan Antonio Osorio <ozz@stacklok.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Juan Antonio Osorio <ozz@stacklok.com>
rdimitrov
left a comment
rdimitrov
left a comment
There was a problem hiding this comment.
I didn't looked deeper at the database model, but overall LGTM
blkt
left a comment
blkt
left a comment
There was a problem hiding this comment.
First review, I've yet to read past the Dependency Graph section, but this should represent most of it.
|
Responding to the discussion around normalizing Good suggestion. I looked into this carefully. The tradeoffs: Normalized junction table gives you efficient per-client queries and relational integrity, but at this scale (dozens of rows, 1-3 client apps per skill) those benefits are marginal. It also introduces complexity around whether skills and groups share the same junction table or have separate ones — since skills associate client_apps per-installation (scope + project) while groups associate them at the group level, a shared table needs qualifiers that get awkward fast. JSONB array maps directly to the Go Going with JSONB for now. If query patterns evolve to justify normalization, it's a straightforward migration. Same pattern will apply to groups when they're migrated. Also renaming the column from |
Revise schema based on reviewer feedback from blkt, rdimitrov, and aponcedeleonch: - Add thin `entries` identity table shared across lifecycle-tracked entities (skills, workloads). Groups are standalone. - Change scope model from user/system to user/project with project_root column for multi-project support. - Use JSONB (BLOB) for tags and client_apps columns, confirmed supported by modernc.org/sqlite (SQLite 3.51.2). - Rename clients to client_apps to align with ClientApp domain type. - Add OCI artifact tracking (reference, tag, digest) from skillet. - Add skill_dependencies and oci_tags tables. - Document cross-platform XDG paths via adrg/xdg library. - Drop premature status index. - Remove resolved open questions (WAL checkpoint, client normalization). - Update future phase tables, alternatives, and references. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
JAORMX
dismissed stale reviews from aponcedeleonch and rdimitrov
via
5b92436
4 participants
Summary
Proposes replacing ToolHive's file-based local state management with a single SQLite database (
toolhive.db).Phase 1 targets the skills subsystem to unblock toolhive#3648. Later phases will migrate workload statuses and groups. RunConfigs and
config.yamlremain file-based.Key Design Decisions
modernc.org/sqlite(pure Go) -- required because ToolHive builds withCGO_ENABLED=0thv serve, CLI commands, and detached proxiespkg/storage/package with domain-specific interfaces (not extending the stream-orientedpkg/state/)toolhive.dbat$XDG_STATE_HOME/toolhive/rather than per-subsystem databasesRelated
🤖 Generated with Claude Code