Skip to content

ServerKit is a lightweight, modern server control panel for managing web applications, databases, and services on your VPS or dedicated server. Built with Python/Flask backend and React frontend.

Note traffic, edge caching and attachments in docs

Juan Denisc81d452f4a59
diff --git a/ROADMAP.md b/ROADMAP.md
index c2ac52c9..188c47e0 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -10,6 +10,10 @@ This document outlines the development roadmap for ServerKit. Features are organ
### Recently Completed (v1.7.x)
+- **See Where an App Is Slow** - Every site nginx serves now logs how long each request took, what the upstream took and whether the cache answered, and the panel rolls that up per app: requests, 4xx/5xx, p50/p95/p99 latency and cache hit ratio on the app's Overview, kept for two days by the minute and a month by the hour. A rule-based check reads those numbers together with the app's and its database's CPU and names the likely bottleneck (a busy app over an idle database, one query dominating a busy database, errors during deploys, a cache that misses), with the fix and that fix's failure mode, and never adds anything on its own. Apps whose containers keep crashing and restarting are flagged with an incident instead of looking "running" between crashes, and each database container gets an Insights tab with its top queries, connections in use and buffer-cache hit ratio.
+- **Faster Sites Without New Moving Parts** - Responses are compressed (gzip, and brotli where the nginx build has it) for every site; the snippet only adds what your nginx.conf doesn't already set, because a repeated directive stops nginx from loading. The page cache became a real one: each site sets its own lifetime, visitors get the last good page while the app errors or refreshes, a stampede of misses reaches the app as one request, and clearing it clears that site only. Proxied sites can long-cache fingerprinted assets such as `app.3f9a2c1d.js`, and the Cloudflare extension adds Cache Rules presets and an opt-in purge of an app's hostnames after every deploy.
+- **Attach a Cache, a Bucket, a Queue or Tracing** - Installing Redis and having an app use Redis used to be two unrelated things. From an app's Overview you now attach a cache, object storage, a queue or tracing: ServerKit reuses an installed service or installs one, writes the connection settings the app's libraries expect (`REDIS_URL`, the `S3_*` and `AWS_*` set, `AMQP_URL`, the standard OpenTelemetry variables), and puts both on a private service network. Storage gives each app its own bucket with a key that can only reach it; the secret stays in the vault. Object storage is now Garage, because MinIO stopped publishing images, and RabbitMQ and an OpenTelemetry Collector join the catalog. Grafana gets Prometheus and Loki as data sources, plus a dashboard of your servers, by attaching them, and an installed Prometheus scrapes the panel with a token the panel issues itself. PostgreSQL can run PgBouncer beside it for apps that opt into the pooled URL, and build-pack apps run their Procfile `worker:` and `scheduler:` lines next to the web process.
+
- **Managed Databases Join the Recycle Bin** - Deleting a managed database no longer destroys it on the spot. Like applications before it, a deleted database now lands in the Recycle Bin with its content and backup policy intact and restorable; only an explicit purge (or asking for the data to be dropped at delete time) is destructive. The distinction that matters: delete keeps your data, purge is the irreversible step, and the UI says which one you are about to do.
- **One Error Shape, With a Request ID You Can Quote** - Every API error now comes back in one consistent JSON envelope carrying a request correlation id, so "it failed" in a bug report can be matched to the exact request in the panel's error monitor. Behind that, crashes a route used to catch and quietly convert into a vague 200 or a bare 500 — the PR-webhook receiver, the job queue bus, and several others — are now recorded in the error monitor every time, with tests pinning that a swallowed crash is a bug, not a behavior.
- **Live Streams and Polling That Behave** - Panel pages pause their background polling when the tab is hidden and catch up the moment you return, so an idle dashboard tab stops costing server load and laptop battery. Live output (deploys today, more run kinds on the same rails next) streams through one event scheme with a shared room grammar, and the deploy console keeps working unchanged through the transition.
@@ -518,8 +522,6 @@ Full DNS record management with provider API integration.
Go beyond basic virtual hosts — full reverse proxy and performance configuration.
-- [x] Visual reverse proxy rule builder (upstream servers, load balancing methods)
-- [x] Load balancing configuration (round-robin, least connections, IP hash)
- [x] Caching rules editor (proxy cache zones, TTLs, cache bypass rules)
- [x] Rate limiting at proxy level (per-IP, per-route)
- [x] Custom location block editor with syntax validation
diff --git a/docs/NEW_FEATURES.md b/docs/NEW_FEATURES.md
index 894fd554..982ee727 100644
--- a/docs/NEW_FEATURES.md
+++ b/docs/NEW_FEATURES.md
@@ -52,10 +52,6 @@ App badge: `app.sleep = {enabled, asleep, idle_timeout_minutes}`.
Idle is measured from `last_activity_at` (bumped on wake / `record_activity`); a
no-activity-baseline policy is never slept blind.
-### Container auto-scale (removed 2026-09-28)
-Removed with the other replica pieces: nginx only proxied to one port and
-the auto-scale sweep was never scheduled. One live copy per app.
-
### GPU monitoring — `services/gpu_service.py` (no model)
| Method | Path | Auth | Notes |
|---|---|---|---|
@@ -100,9 +96,9 @@ the auto-scale sweep was never scheduled. One live copy per app.
## 5. Operational notes
-- **Cron the sweeps**: `POST /apps/sweep-idle` (auto-sleep) and `POST /apps/scale-sweep`
- (auto-scale) are admin endpoints meant to be hit periodically. Wiring them to a
- built-in scheduler is a follow-up; for now drive them from cron.
+- **Cron the sweep**: `POST /apps/sweep-idle` (auto-sleep) is an admin endpoint meant
+ to be hit periodically. Wiring it to a built-in scheduler is a follow-up; for now
+ drive it from cron.
- **Public DDNS endpoint**: `/ddns/update` is the only unauthenticated route added
(the per-host token is the credential). Serve it over HTTPS.
- **WAF integration is additive**: per-app rules go to `serverkit-conf.d/waf` and an