mbs-api ran single-worker uvicorn with no --workers flag at all — one
event loop handling every request. Now install.sh (and mbs update, so
existing installs pick it up too) compute a worker count from nproc
(clamped 1-4, matching typical VPS core counts) and bake it into the
systemd unit via sed substitution of a __WORKERS__ placeholder. Safe
to parallelize: verified no api.py module-level mutable state, all
of it already goes through sqlite (payment idempotency and the
xray-config file lock are already correct across separate processes,
not just asyncio tasks within one — confirmed both are OS/db-level,
not in-process). Verified with a mock dry-run of the new systemd-unit
section (fake nproc, real sed substitution) producing the expected
ExecStart line for several core counts.
Also: build_subscription_text() — called on every single hit of
/sub/{token}, the single most frequently called endpoint in the whole
app, since every VPN client re-fetches on every reconnect — was doing
one db.get_node() call per distinct node in a user's subscriptions
instead of fetching once. Same N+1 shape as the admin-endpoint bugs
fixed yesterday, except this one is on the hot path, not just the
admin panel. Fixed to batch-fetch via db.list_nodes() once.
Verified: correct output for a 4-node subscription (each node's
address appears exactly once, de1's 4 transports all present,
nothing silently dropped) and measured ~1.3ms/call average.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
install.sh clones from the mirror first (falls back to github.com if
unreachable); mbs update fetches origin (mirror) first, falls back to
a github remote if that fetch fails. Mirror itself is a bare repo on
финка2, kept in sync from GitHub every 10 min via an authenticated
token (needed because that box's IP gets rate-limited/blocked by
GitHub for anonymous git clones).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
