The November 14, 2026 investigation found two separate costs: the configured public NEAR RPC took about three seconds per view from the VPS, and the board waited for three views in sequence. GitHub descriptions could add further waits.
Read-only measurements and their context are recorded in PERFORMANCE_MEASUREMENTS.json. No transactions were sent while profiling.
| Measurement | Before | After |
|---|---|---|
| Public board API, successful sample | 9,993 ms | 1,733 ms first observed / 222 ms repeated, after the provider change alone |
| Public leaderboard API, successful sample | 3,267 ms | 1,167 ms first observed / 265 ms repeated, after the provider change alone |
| RPC status + finalized view, from the VPS | Official endpoint: 3,175 / 3,052 ms | FastNEAR: 83 / 53 ms; dRPC: 128 / 65 ms |
| Revised local board read model | — | 567 ms empty read cache / under 1 ms cached |
| Revised local single-bounty detail | — | 204 ms empty read cache / under 1 ms cached |
| Revised local leaderboard | — | 205 ms empty read cache / under 1 ms cached |
| Revised local contributor profile | — | 279 ms empty read cache / under 1 ms cached |
These are individual diagnostic samples, not percentiles or a load test. Public API timings include network/TLS and HTTP overhead; the first post-change configuration request alone took 1,146 ms versus 222 ms on repeat. Local revised numbers measure exported server data functions using real finalized testnet state, an empty metadata database, and explicit read-cache invalidation before each screen. They do not measure production HTTP or browser rendering. The warm transport is reused after the first local query. The raw JSON distinguishes all test conditions.
Initial public board/detail/leaderboard/profile samples returned 503 before the successful baseline samples. The app exposed the failure honestly; it did not replace the unavailable chain state with demo balances. A provider improvement does not guarantee future testnet availability.
The NEAR RPC provider documentation
lists https://test.rpc.fastnear.com and https://near-testnet.drpc.org as public
testnet endpoints. Read-only checks from the actual web container confirmed
chain_id: testnet and current finalized contract views on both. The operator
changed the hosted NEAR_RPC_URL to FastNEAR after those checks. The shared public
configuration now uses FastNEAR as its default; an explicit environment value
still takes precedence.
The application does not silently switch RPC providers during transaction
submission, re-sign a transaction, or add a broadcast retry. Public contract
views use the documented query / call_function JSON-RPC shape with
finality: final, strict response decoding, no redirects, and a six-second
abort deadline. This avoids near-api-js 7.5.3's default twelve HTTP attempts
holding a public page open. The verified SDK remains responsible for NEP-413,
access-key checks, transaction status, and the separate signing services.
- The board starts contract configuration and the first bounty page together. It does not request configuration twice. Additional pages run at most four at a time, and incomplete/noncontiguous rows fail rather than silently dropping escrow records. A first page that raced new funding is retried once.
- Detail reads request only the selected bounty and configuration. Contributor profile data starts alongside its board read instead of after it.
- A process-wide, bounded display cache coalesces concurrent reads for five
seconds. A slow initial read remains shared while pending; its successful
result gets a usable cache lifetime. Composite responses retain older
updatedAtboundaries instead of extending them. Namespace keys include mode, contract account, and RPC URL. - Confirmed transaction responses invalidate display caches. In-flight older
requests cannot refill an invalidated cache.
getBountyandviewContractremain uncached authority reads for confirmation and appeal eligibility. - GitHub titles/descriptions no longer block escrow responses. Known text is
served immediately while stale/missing descriptions refresh with three workers,
four-second timeouts, and retry backoff. Failed refreshes do not overwrite good
text with placeholders. Cold instances display the real issue number/link
until descriptive text arrives.
metadataUpdatedAtandmetadataStalemake that distinction explicit. - Public board/detail/profile/leaderboard responses expose
updatedAt,Server-Timing,X-GitSettle-Read-At, and the five-second cache TTL. HTTP responses remainCache-Control: no-store. Client display caching must retain that read timestamp and clear on confirmed wallet updates.
Descriptive metadata can be stale. Monetary display snapshots are short-lived finalized observations; failed refreshes return an error, with no stale-success fallback and no illustrative/live data mixing. A cached screen never authorizes a transfer: the contract and uncached transaction confirmation remain decisive.
For comparable public HTTP samples before and after a deployment, use the same machine and this read-only script:
npx tsx scripts/measure-performance.ts \
--origin https://gitsettle.89.167.109.196.sslip.io \
--output /tmp/gitsettle-api-after.jsonIt makes five sequential rounds across configuration, board, bounty #8, and
leaderboard. The report includes each HTTP status, total/header time, parsed
Server-Timing durations, medians, and allowlisted counts/state summaries. It
contains no response bodies, cookies, credentials, or signatures. The origin can
also come from the exported APP_ORIGIN; .env files are not loaded. Omitting
--output prints the report without creating a file. Existing artifacts are
never overwritten. Missing server timing metrics remain empty rather than being
invented. --timeout-ms defaults to 15,000; use --help for options.
Ten focused regressions cover concurrent board reads, blocked GitHub metadata, direct detail loading, uncached authority reads, parallel profile fetches, RPC namespace isolation, bounded/complete pagination, malformed RPC responses, success-only cache expiry, failure recovery, confirmation invalidation, and composite snapshot age. Run:
npx tsx --test packages/shared/src/read-cache.test.ts packages/shared/src/data.test.ts
npm run typecheckThe source changes passed those checks. Production build and browser/layout validation are coordinated with the accompanying UI revision; the measurement file does not claim that the revised source was already deployed when sampled.
The public-view navigation cache is bounded to 100 URLs, coalesces in-flight requests and respects the server snapshot timestamp. It excludes all private/auth/AI endpoints. A URL change never displays another account's previous record, and confirmed transactions invalidate pending and cached observations. Mounted content remains stable during a refresh, hidden tabs do not poll, and visible tabs resume with a fresh read. Six isolated regressions cover the cache boundary and invalidation races.
The custom wallet picker opens without loading signing SDKs. Choosing a provider or restoring an existing connection initializes the required wallet code. Persisted transaction recovery runs before optional chunk loading, so a failed download cannot lose a sent request. Mismatched return evidence preserves original intent for recovery; only a verified matching finalized failure permits clearing it. Browser checks cover provider selection, deferred SDK loading and failure/recovery paths.
Runtime source a63ff25c6d69733ce253da15546c22b3142150c2 was verified over public HTTPS on November 14, 2026. PERFORMANCE_LIVE.json contains five samples per endpoint: all 20 returned HTTP 200. Median board response was 251 ms, detail 233 ms, and leaderboard 225 ms. The first measured board call took 297 ms, including 73.6 ms of server read-model work; later samples recorded 0.1–0.2 ms of server work. This sample does not force a process cold start.
BROWSER_PERFORMANCE.json records a fresh Chromium visit to this source: the first bounty row appeared after 2.142 seconds, and 15 JavaScript requests transferred 203,428 bytes. Opening the actual provider picker in a separate signed-out page added no JavaScript requests; no provider was selected. Full loading includes DNS/TLS, assets, hydration and the visitor’s connection. This is one diagnostic observation, not a latency percentile or a statistically controlled comparison.
The earlier premium source 85c4c3bab27f37d1fce19d5987d2f8913e88a42c is preserved in PERFORMANCE_LIVE_85C4C3.json and BROWSER_PERFORMANCE_85C4C3.json: board/detail/leaderboard medians were 222/221/225 ms, and one browser visit reached its first row in 3.092 seconds with 16 JavaScript requests and 199,312 transferred bytes. The original-UI observation after the RPC switch was 3.179 seconds, 21 JavaScript requests and 419,780 bytes. These observations record their respective sources and network conditions; they do not establish a measured latency improvement between UI releases.