hetzner-radar/src/ at main · else42.tngl.sh/hetzner-radar · Tangled

else42.tngl.sh / hetzner-radar

0

[READ-ONLY] Mirror of https://github.com/elsbrock/hetzner-radar. 🕵️Track prices of the Hetzner dedicated server auction radar.iodev.org/

cloudflare cloudflare-d1 cloudflare-workers duckdb hetzner sqlite sveltekit tailwind wasm

0

hetzner-radar / src /

at main 3 folders 6 files

Claude fix(cloud-status): stop availability history contradicting the live table 5d ago

fix(cloud-status): stop availability history contradicting the live table The availability heatmap reported pairs as available while the status table — reading the live snapshot — correctly showed them unavailable (#286). `cloud_availability_v2` stores availability *transitions*, not samples, but the history query aggregated them into hourly buckets with `MAX(double1)`. Any pair that flipped twice inside one hour lost its "became unavailable" edge, and since the client holds the last known state until the next transition, a few minutes of stock rendered as available ever since. This hit scarce, flapping server types hardest — exactly the ones being watched. Query raw transitions at full resolution instead, and resolve the state entering the window with a second `argMax` query over the preceding 30 days rather than inferring it by inverting the window's first edge — a heuristic that only holds while edges strictly alternate, which bucketing had broken. The trailing bucket of a live window is now reconciled against the snapshot the table renders, but only when it is saturated in the opposite direction: that can only mean a transition never reached the dataset, whereas a partly available bucket is consistent with either state and stays as measured. Extract the step-function reconstruction into $lib/availability-history so the seeding and bucketing rules are covered by tests.

5 days ago

fix(cloud-status): stop availability history contradicting the live table The availability heatmap reported pairs as available while the status table — reading the live snapshot — correctly showed them unavailable (#286). `cloud_availability_v2` stores availability *transitions*, not samples, but the history query aggregated them into hourly buckets with `MAX(double1)`. Any pair that flipped twice inside one hour lost its "became unavailable" edge, and since the client holds the last known state until the next transition, a few minutes of stock rendered as available ever since. This hit scarce, flapping server types hardest — exactly the ones being watched. Query raw transitions at full resolution instead, and resolve the state entering the window with a second `argMax` query over the preceding 30 days rather than inferring it by inverting the window's first edge — a heuristic that only holds while edges strictly alternate, which bucketing had broken. The trailing bucket of a live window is now reconciled against the snapshot the table renders, but only when it is saturated in the opposite direction: that can only mean a transition never reached the dataset, whereas a partly available bucket is consistent with either state and stays as measured. Extract the step-function reconstruction into $lib/availability-history so the seeding and bucketing rules are covered by tests.

5 days ago

perf: lazy-load DuckDB, fix FontAwesome and banner CLS, slim carousel - DuckDB WASM + database no longer load on every page; initialized on demand by analyze/configurations/statistics/server drawer/SQL console, with the library dynamically imported so it stays out of the common bundle - sb.duckdb cache-buster bucketed to 5-minute windows (was Math.random(), forcing a full re-download per page view); explicit reloads stay fresh - FontAwesome CSS shipped statically (autoAddCss=false) - fixes nav icon/label reflow on load - banner is SSR-visible and dismissed state is applied pre-paint via inline script, eliminating the pop-in layout shift - screenshot carousel: no more eager preload of all 10 PNGs (~3.4 MB), lazy loading, explicit dimensions to reserve space - hero card stack rotation paused below lg where the stack is hidden

1 week ago

fix(cloud-status): stop availability history contradicting the live table The availability heatmap reported pairs as available while the status table — reading the live snapshot — correctly showed them unavailable (#286). `cloud_availability_v2` stores availability *transitions*, not samples, but the history query aggregated them into hourly buckets with `MAX(double1)`. Any pair that flipped twice inside one hour lost its "became unavailable" edge, and since the client holds the last known state until the next transition, a few minutes of stock rendered as available ever since. This hit scarce, flapping server types hardest — exactly the ones being watched. Query raw transitions at full resolution instead, and resolve the state entering the window with a second `argMax` query over the preceding 30 days rather than inferring it by inverting the window's first edge — a heuristic that only holds while edges strictly alternate, which bucketing had broken. The trailing bucket of a live window is now reconciled against the snapshot the table renders, but only when it is saturated in the opposite direction: that can only mean a transition never reached the dataset, whereas a partly available bucket is consistent with either state and stays as measured. Extract the step-function reconstruction into $lib/availability-history so the seeding and bucketing rules are covered by tests.

5 days ago

feat(mcp): public MCP server with authenticated alert management POST /mcp speaks stateless Streamable HTTP JSON-RPC. Read tools work with no credentials; presenting a valid MCP access token additionally exposes alert management, so tools/list is conditional on the Authorization header. - search_auctions / get_auction serve from the KV snapshot, so MCP traffic reads zero D1 rows. cloud_availability proxies the worker. - list_alerts / create_alert / delete_alert are gated on a Better Auth MCP session. Migration 0017 adds the OAuth provider tables (camelCase, unlike the rest of the schema: MCPOptions has no field-mapping escape hatch and nothing in the app queries them). - Alert tools take the same flat schema as search_auctions and build the ServerFilter server-side, in defaultFilter's key order — idx_price_alert_ user_id_filter is UNIQUE on the serialised string, so key order is load-bearing. Encodings were taken from MATCH_ALERTS_SQL rather than the UI, and verified against production: RAM is log2(GB), disk sizes are units of 500 GB, and price_alert.price is GROSS whole euros with vat_rate as a percentage — not cents and not a decimal rate, as the spec had assumed. Also fixes baseURL, which was pinned to the production origin. Better Auth's isAuthPath() rejects a mismatched origin, so every /api/auth/* route 404'd on localhost and would have on any preview deployment.

1 week ago

perf: lazy-load DuckDB, fix FontAwesome and banner CLS, slim carousel - DuckDB WASM + database no longer load on every page; initialized on demand by analyze/configurations/statistics/server drawer/SQL console, with the library dynamically imported so it stays out of the common bundle - sb.duckdb cache-buster bucketed to 5-minute windows (was Math.random(), forcing a full re-download per page view); explicit reloads stay fresh - FontAwesome CSS shipped statically (autoAddCss=false) - fixes nav icon/label reflow on load - banner is SSR-visible and dismissed state is applied pre-paint via inline script, eliminating the pop-in layout shift - screenshot carousel: no more eager preload of all 10 PNGs (~3.4 MB), lazy loading, explicit dimensions to reserve space - hero card stack rotation paused below lg where the stack is hidden

1 week ago

fix: update Playwright tests to match current UI content - Update landing page tests to match actual headings - "Powerful Features for Smart Server Hunting" instead of "Key Features" - "Your Path to Server Success" instead of "How It Works" - Increase tolerance in advanced filtering tests from 10 to 20 to account for data variability - Format all files with prettier All 53 Playwright tests now pass successfully. 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <noreply@anthropic.com>

1 year ago

hooks.server.ts

fix(mcp): require explicit consent, and stop building inverted filter ranges Two fixes. 1. Consent. Better Auth's mcp plugin only asks for consent when the request carries prompt=consent (requireConsent: query.prompt === "consent"), otherwise it issues an authorization code the instant a signed-in user hits /authorize. Clients do not send it. Combined with open client registration that let anyone register a client pointing at their own redirect_uri, lure a signed-in user to a URL, and receive a working code with no interaction — PKCE is no defence when the attacker is the client. forceConsentHandle now injects prompt=consent, and /auth/consent shows who is asking and what for. The client name comes from open registration, so it is presented as unverified. 2. Inverted ranges. A minimum above the UI slider's default ceiling produced e.g. hddInternalSize [128, 44] for a 64 TB request, and MATCH_ALERTS_SQL then evaluates `hdd_size >= 64000 AND hdd_size <= 22000` — never true, so the alert silently never fired. The ceiling is now raised to at least the floor. Caught by a spot-check of a filter echoed back from create_alert. Consent verified end to end: authorize -> prompt injected -> consent page -> approve -> callback with code and state intact.

1 week ago

fix: update Playwright tests to match current UI content - Update landing page tests to match actual headings - "Powerful Features for Smart Server Hunting" instead of "Key Features" - "Your Path to Server Success" instead of "How It Works" - Increase tolerance in advanced filtering tests from 10 to 20 to account for data variability - Format all files with prettier All 53 Playwright tests now pass successfully. 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <noreply@anthropic.com>

1 year ago