Xiantao Cai | Building Connexa

Building Connexa for multilingual teams. React/TypeScript engineer sharing product decisions, bugs, and experiments in public.

[Xiantao Cai | Building Connexa @caixiantao]

An invite link can fail without expiring.

I fixed a Connexa bug where auth erased the original team invite:

invite → auth → return → join

I also added safe return-path validation and completion-based events.

Small redirect. Entire collaboration moment.

[Xiantao Cai | Building Connexa @caixiantao]

4/ 全是公开数据源，不用付费 key。不构成投资建议，只是把账摆出来：
https://t.co/4P7c6TLWJm

#美股 #纳指 #QDII

[Xiantao Cai | Building Connexa @caixiantao]

1/ 想买纳指的朋友注意：最近不少场外纳指基金把单日申购限到 5–100 元，有的直接暂停。大家只好转去场内买 ETF，可 9 月 30 日场内纳指 ETF 普遍比参考净值（IOPV）贵 10% 以上。花 1 万，相当于多付 900 多块。

[Xiantao Cai | Building Connexa @caixiantao]

3/ 最在意的一条：不拿不同时间点的数凑。溢价只用同一次行情返回的价格和 IOPV 算，还要和数据源自己给的溢价对一遍，对不上就不显示。接口挂了，就显示上次成功的值和它原本的日期。

[Xiantao Cai | Building Connexa @caixiantao]

4/ Every number carries when it was fetched and where it came from. Public sources only, no paid keys, shared server cache.

Not investment advice, just the receipts:
https://t.co/4P7c6TLWJm

[Xiantao Cai | Building Connexa @caixiantao]

1/ Want Nasdaq-100 exposure from mainland China? Many onshore index funds have cut daily purchases to ¥5–100 or paused them. So demand moves to exchange-traded Nasdaq ETFs — and on Sep 30 those traded 10–11% above their indicative NAV (IOPV).

[Xiantao Cai | Building Connexa @caixiantao]

3/ The rule I cared about most: never mix time points. Premium is computed only when price and IOPV come from the same quote response. If a source fails, the page shows the last good value with its original date, not a fresh-looking guess.

[Xiantao Cai | Building Connexa @caixiantao]

@0xAnil404 Easy to reproduce, too. My compose block set no stdin_open and no tty, so the stdio MCP server read EOF the moment it started and exited 0. restart: unless-stopped did the rest. A test that asserts on restart count, not exit code, would catch it.

[Xiantao Cai | Building Connexa @caixiantao]

@tanujDE3180 The case that bit me: a Docker healthcheck running through Node 18+ hit localhost, which resolved to ::1 first. The app was bound to IPv4 only, so ECONNREFUSED. The container was marked unhealthy for 4 weeks while it was serving fine.

[Xiantao Cai | Building Connexa @caixiantao]

@helltar_ The file-age check also covers the opposite failure, which I hit: a healthcheck that was itself wrong. One probed :5001 while the service listened on 5100, another hit ::1 against an IPv4-only bind. Both containers served fine the whole time they were marked unhealthy.

[Xiantao Cai | Building Connexa @caixiantao]

@acecreamu @fromcodetocloud Honestly, no alert caught it. I found it during a manual audit of the box: docker inspect showed RestartCount 39,352 with ExitCode 0 and OOMKilled false. That combination gave it away: clean exits, no memory kill, still restarting.

[Xiantao Cai | Building Connexa @caixiantao]

@syssignals Agreed. The habit that would have caught both of mine: run the check inside the container once before shipping it. One probed :5001 while the service listened on 5100, the other hit ::1. Both would have failed on the very first run.

[Xiantao Cai | Building Connexa @caixiantao]

@PratikPatel_227 One gotcha once cloudflared runs in Docker: localhost becomes the cloudflared container, not your machine. I had 3 routes on the bridge gateway and 1 on localhost. That one 502'd while the others worked. Use the service name (or host.docker.internal on Docker Desktop).

[Xiantao Cai | Building Connexa @caixiantao]

Full write-up: https://t.co/y5MXveHYfK

[Xiantao Cai | Building Connexa @caixiantao]

I exported 11 years of my notes. 493 of the 976 short ones contain no period at all. Just nouns.

They weren't notes. They were a cache with no retention policy.

"put it down gently / go to Dali" (2015): the pointer survived. Whatever it pointed to didn't.

[Xiantao Cai | Building Connexa @caixiantao]

The lesson I kept:

Restart count is a better smoke alarm than exit code. Exit 0 only tells you the process agreed to leave — not that it should have.

Full write-up (the stdin EOF loop + the IPv6 localhost trap): https://t.co/6Gs4C5JYI9

[Xiantao Cai | Building Connexa @caixiantao]

A container on my server had restarted 39,352 times. Every exit code was 0.

That's why nobody noticed. Monitoring watches for crashes. This wasn't a crash — the process started, found nothing to do, exited 0, and restart: unless-stopped brought it back. 🧵

[Xiantao Cai | Building Connexa @caixiantao]

So on one box, at once:
• two containers screaming "unhealthy" that were serving traffic fine
• one container calmly exiting 0 that was broken

The health signal was inverted in both directions. Trust it and you'd fix the healthy ones and never see the broken one.