Tag: history

Reddit 主要是用 Python 寫的，這邊介紹了他們歷年來的 Code Deploy 系統：「The Evolution of Code Deploys at Reddit」。

最早期的時候 (2007 到 2010) 是用 rsync 更新程式碼，然後跑個迴圈用 ssh 連進去重跑：

# build the static files and put them on the static server
`make -C /home/reddit/reddit static`
`rsync /home/reddit/reddit/static public:/var/www/`

# iterate through the app servers and update their copy
# of the code, restarting once done.
foreach $h (@hostlist) {
    `git push $h:/home/reddit/reddit master`
    `ssh $h make -C /home/reddit/reddit`
    `ssh $h /bin/restart-reddit.sh`
}

2011 時因為人變多了，用 IRC 把過程丟出來 (okay，我知道你想問的問題... Slack 是 2013 年推出的)：

The process for actually doing the deploy looked the same, but now the system did the work for you and told everyone what you were doing.

另外值得一提的是，因為他們不是自己架 IRC server 而是用外面第三方的伺服器，所以他們決定 IRC 只有單向告知的功能：

There was a lot of talk of systems that managed deploys from chat around this time, but since we used third party IRC servers we weren’t able to fully trust the chat room with production control and so it remained a one-way flow of information.

2012 時則是把機器列表放到 DNS 上，某種 service discovery 系統：

First, it fetched its list of hosts from DNS rather than keeping it hard-coded. This allowed us to update the list of hosts without having to remember to update the deploy tool as well — a rudimentary service discovery system.

另外是固定的版本，而非拉 master 下來，這樣可以避免 race condition 的不一致性 (推到一半有人把 code 塞進 master)：

Another small but important change was to always deploy a fixed version of the code. The previous version of the tool would update master on a given host, but what if master changed mid-deploy because someone accidentally pushed up code? By deploying a specific git revision instead of branch name, we ensured that the deploy got the same version everywhere in production.

2013 往雲上搬，於是遇到像是「開新機器時剛好在 deploy 會拉到舊的 code」這種 edge case。

What happens if a server is launched while a deploy is ongoing? We had to make sure each newly launched server checked in to get new code if present. What about servers going away mid-deploy? The tool had to be made smarter to detect when the server was gone legitimately rather than there being an issue with the deploy process itself that should be noisily alerted on.

2014 遇到機器數量太多，推一輪要一個小時而被迫要平行化處理：

Over time, the number of servers needed to serve peak traffic grew. This meant that deploys took longer and longer. At its worst, a normal deploy took close to an hour. This was not good.

2015 則是加上 deploy lock，避免同時間有兩個人在 deploy：

Engineers would ask for the deploy lock and either get it or get put in the queue. This helped keep order in deploys and let people relax a bit while waiting for the lock.

2017 的部份則是提到了伺服器的數量：

This new mechanism allows us to deploy to a lot more machines concurrently, and deploy timings are down to 7 minutes for around 800 servers despite the extra waiting for safety.

看起來到現在還是維持手動 deploy，而不是自動化... 這塊還蠻有趣的 :o