Joel отказался от fb и tw
https://www.joelonsoftware.com/2018/01/12/birdcage-liners/
- Stream is always full of conflict and stress. Gave up, stopped using them.
- As he designs social software for a living, he should have a professional opinion why Tw and Fb made him unhappy
- Tw: I care about friends, I keep in touch with them.
- Everybody don’t use twitter to share happy news, only for some flame wars.
- 140 symbols is not enough: You start with complicated thought, then shorten it, the result - is very trivial idea without nuances
- Then it is incorrectly comprehended, then the reply is incorrectly read. Then you’re accused of smth non-related.
- The problem of Tw: we have problems of explaining emotion in text, and in Tw we shorten the text. That’s the second level.
- Outrage and indignation - what makes it work. Not original idea -> selection. Very attractive. 140 - is weird but work well.
- Fb: copied Tw idea of “feed” to keep you coming back compulsively.
- Added ML: we are not going to show everybody every post. We will choose what is important (to come you back)
- The only thing ML was told to care - addiction (“engagement”)
- Dog addiction: dog is not always given the treat. Only sometimes. To obey even without treat.
- FB doesn’t always show smth important. A lot of junk I don’t need.
- But I need to come back to get important sometimes.
- Both Tw and Fb’s selfish algorithm are optimized to increase the number of hours we spend there.
- Researcher discovered that algorithms served to divide the world into smaller groups. United by some interest.
- Political party. You’ll get news you agree with or agree on blaming.
- You ‘LIKED’ or ‘RETWEETED’ and feel like you had done smth important. But you hadn’t.
- Lesson: when you design software - you create the future.
- Decision to limit length or to use ML to maximize engagement - will have social impact that is difficult to predict.
- We should make the world better. Or not worth than before. If invention spins out of control - we have to fix it.
- Good news: Facebook announced that they change the direction. Will want the feed to leave people feeling “more connected and less lonely”
- “sacrifice engagement”: reduce the time spent on Facebook, but the time will be more valuable.
- Facebook has grown up?
Патчи против Meltdown & Spectre в Linux - полный мусор
https://habrahabr.ru/company/cloud4y/blog/347418/
- Интел предупреждает, что не надо устанавливать обновления CVE-2017-5715, пока интел не разработает “нормальное” решение
- Предполагается, что пользователи будут во время загрузки выбирать: производительность против безопасности (кнопку нажать?)
- Три новых аппаратных патча (для микрокода CPU):
- Indirect Branch Restricted Speculation (IBRS)
- Indirect Branch Prediction Barriers (IBPB)
- Page Table Isolation (PTI)
- Linus about patches
Scaling Kubernetes to 2500 Nodes
Из тем слушателей
https://blog.openai.com/scaling-kubernetes-to-2500-nodes/
- Deep learning research on k8s - for 2 years.
- Several clusters, one of which is 2500nodes. Azure, combination of D15v2 and NC24 VMs
- Issues:
- etcd: kubectl had regular timeouts. New master nodes don’t help. Found: write latency spiking to hundreds of ms.
moved etcd to work with local (not network) ssd
monitoring overload the apiservers, reduced the polling level
- tweak: store kubernetes events on a separate etcd cluster. (dataflow / controlflow split)
- kubedns antiaffinity rules
- Dota project - 17GB image. “Pending state” problem.
TO fix: serialize-image-pulls -> false (requires overlay2 instead of AUFS)
image-pull-progress-deadline to 30min