Comments by "Megakoresh" (@gigakoresh) on "ByteByteGo" channel.

Don't forget to also add instrumentation to your app when possible. A lot of time will be saved when debugging a properly instrumented system, because there is no need to add all those logs and wait for the customer to trigger the bug again. You can prove your theory by looking at the function calls.
18
Another excellent video. Could you plz share how you create these animated diagrams? What software do you use?
16
Inspiring story. It would be really cool to see how Telegram solves their backend challenges. They have an even greater scale, smaller team with less funding and fastest latency of any messenger. I wish they shared more of their backend, that system is engineering marvel, just like Discord.
7
Amazing content, I am gonna recommend these books be bought by the company and put to premises where people can read them to understand these concepts. I have a suggestion for a video btw: can you explain CPU metrics and throttling on container systems? There is lots of misunderstanding especially with Kubernetes resources like "millicores" about what it actually means and how the usage is calculated and when throttling is performed on Kubernetes nodes and how. Also what are the implications of these metrics and quotas on CPUs with many cores and the difference between nodes running on hypervisor vs bare metal. It's a hole can of worms of course, but I think just starting with the quotas and throttling would already clear things up a lot.
4
Really good content, as always. You can also add a Kafka or some other streaming service to the business service, so that writes send events to the streaming system. You can then connect that streaming system to the write database as a sink, which will allow you to distribute the write loads evenly throughout the day and handle peak loads without consuming extra resources. Streaming systems like Kafka are so heavily optimized for such use-cases that even a relatively smol cluster with 3 4Gb+2CPU+100Gb should easily be able to handle these loads and have a lot of headroom if you use z-compression on the topic. As an added benefit you can perform change management on the database transparently to the user because streaming system will buffer all write operations automatically while the sink is offline. Streaming systems generally pair with a heavy read/low write system quite well.
4
@andyl.5998 If you ask instead of googling it's probably not yet time for you to learn about that haha. But for sake of completeness - it's a framework and a cloud service that collects detailed metrics about your application runtime for inspection. Things like exceptions, function calls, heap and stack sizes, session count and so on.
3
@andyl.5998 My point was that whatever I explain here in the YT comment isn't really gonna help you understand this thing. I did not intend to come off as arrogant or anything, sorry. And if you actually felt like my explanation was helpful, and are not just saying this to be nice, then I am glad to be wrong.
3
I think Docker's biggest mistake was treating Kubernetes as a competitor. They contributed so much to the ecosystem, it was only fair for them to take back. If they jumped on the bandwagon fast enough instead of trying to make Docker Swarm relevant, and produced their own Kubernetes as a Service, I am sure they would have the best one - they have the expertise. But they didn't and now the market is too saturated with Kubernetes distros and services for them to make an entry. So it's not a case of picking a wrong strategy with Open Source, its more a case of not following through fully with your picked strategy. Kubernetes was open from the start. They should have just made their own "Dockernetes" and sold it as managed service and enterprise Kubernetes distro.
1
@minhquando100 I don't want to give bad advice, so no. Instrumentation comes in all shapes and sizes and there is no single answer for what is good. Fot example each cloud provides it's own instrumentation and there are cloud-agnostic instrumentation. There is frontend instrumentation and backend. If your code is simple and not highly performance-dependent instrumentation would only add unnecessary maintenance overhead. When it comes to system design you must do your own research to find what works best for your use-case. All I can do is provide a keyword for a field of technologies to put on your radar if it wasn't there already. Please do not build systems based on some youtube advice.
1
Kafka is way overkill for that. Learned that the hard way. Only makes sense if you already have to maintain a kafka cluster or buy it as a service for something actually data-intensive. Even then, by itself Kafka is a huge pain the ass to configure and maintain. You need a management system on top of it like Confluent or Aiven, if you dont want to fiddle with its terrible CLI tools and build ACLs all day. And those cost an arm and a leg. Also Kafka rarely breaks, but when it does, it's a catastrophe. There is no good way to quickly and automatically recover a broken cluster, and it's very easy to fuck up. I can give you one example - we used a 3-DC cluster where all communication between clients as well as brokers to each other was handled by IP lists. DNS was never used. Then at some point the DNS proxy we used got busted and cluster ended up in split brain state. It was completely incomprehensible why. We had to do network packet monitoring, when we found out that even though the broker didn't use a DNS, it still had a call stack failing silently when contacting brokers on outgoing calls. So all messages coming to the broker were going fine, but all messages going out were failing after a failed DNS lookup for... something that the broker wasn't even gonna use! So other brokers thought the broker was alive and well, and the broken broker thought it was broke (i mean alone, sorry, couldn't resist), constantly trying to assign itself master since it thought the rest of the cluster was dead. Took a whole day to fix and you can imagine the damage since the whole cluster was affected... And this is far from the only example. My experience with Kafka is actually that it's distributed nature is a tool for scalability, not reliability. As far as reliability goes, it does the opposite, each new broker adds more ways for the whole cluster to break. So yeah, Kafka is powerful, about as much as a bulldozer is powerful. If you only want go to shop and back, maybe just take a regular car instead.
1
Squashing doesnt actually destroy the squashed commits, you can still access them by the hash. I use all of these depending on the situation. Squash is good if there are lots of small inconsequential changes, e.g. when debugging CI system. Rebase is good for example for cases when a teammate has updated dependencies and you are adding your feature on top of the new dependencies: merge can be error prone in this case as auto-merge settings can sometimes mess up the syntax. And in all other cases - merge is the way to go. Github, Gitlab and Bitbucket also favor this approach because their branch protection settings typically center around squash and merge, which are better suite for everyday actions that can be done automatically.
1
@aniketshukla9568 Not with that kind of scale, no.
1