When should you shard your database?
VloĆŸit
- Äas pĆidĂĄn 6. 06. 2024
- Database Application level sharding is the process of splitting a table into multiple database instances in order to distribute the load. However, Sharding a database is an expensive operation (maintainability and overhead) and I suggest you do that only when you absolutely need to. That means when your single instance database can no longer serve queries with minimum latency. So I suggest you monitor that and only shard if necessary.. I rather do replication make master / backup and make requests read from replica than Sharding just because it is easier.. good question... nice idea for a video
Time codes
0:00 Intro
5:00 Partition horizontally
7:45 Replication (Master/backup) Scale reads
11:00 Scale writes by region
12:30 Sharding
16:40 Sharding in CZcams
đ Software Architecture Videos
âą Software Architecture
đŸ Database Engineering Videos
âą Database Engineering
đ° Network Engineering Videos
âą Network Engineering
đ° Load Balancing and Proxies Videos
âą Proxies
đ Postgres Videos
âą PostgresSQL
đąDocker
âą Docker
𧟠Programming Pattern Videos
âą Programming Patterns
đĄ Web Security Videos
âą Web Security
đŠ HTTP Videos
âą HTTP
đ Python Videos
âą Python by Example
đ Javascript Videos
âą Javascript by Example
đŸDiscord Server / discord
Support me on PayPal
bit.ly/33ENps4
Become a Patreon
/ hnasr
Stay Awesome,
Hussein - VÄda a technologie
Thanks for the excellent content. I've summarized this for my own understanding.
Before you Shard, try the following first
0. Understand what your actual problem is before optimizing(too slow reads vs too slow writes) Analyze your slowest queriers and see why its slow: czcams.com/video/-qNSXK7s7_w/video.html Create indexes on appropriate columns and tune your data schema.
1. Horizontal Partitioning - Have partition key(mostly on primary key) and split database into different ranges. This will create smaller B-trees on the indexes.
2. Vertical Partitioning - When you have columns that you rarely access, and you cut a column out of the main database. This will make reads faster for frequent queries and slower for not frequent queries and make your B-trees smaller(less space in memory also)
Partitioning Video: czcams.com/video/QA25cMWp9Tk/video.html
----
Thanks : )
Thanks manïŒ
thanks! saved to notes
Thanks dude
Appreciated it
Hussein, you have a great style of presentation with proper tone, cadence, painting pictures without use of displays, and humor thrown in the right places. All this with no sense of arrogance exuded. Love your videos.
Most underrated channel on youtube. Sometimes I get bored from work and I come here to learn something interesting. You don't even know how much your videos mean to me. Thanks a lot Hussein!
I am gonna use social distancing analogy a lot from now on đ
One thing you don't mention which is important to understand is that read replicas cause some load on your primary for replication. Much like anything else, as you said, it isn't free. But it is fairly cheap.
One use case you missed for sharding is data sovereignty. Sometimes data can be split into groups by location (e.g. zip country code) which not only can help with performance but can meet legal requirements for where data lives.
I really love to watch your videos even though those topics are not the part of the my job but I watch it because I know I'll definitely learn something new. So keep updating us and keep uploading, appreciable.
Hi thanks for your efforts. Its very hard to find experienced devs sharing their industrial challenges and providing good insights. Hoping to learn more. Keep up the work. Fighting !âșđđ
I would say under rated channel. He is not teaching but he is discussing in a very informative way
You are such a genius !!!
You started by answering the question very early in the video, I like it.
Then you started explaining everything very nicely !!!
Perfect !!! Go on maaaaaaan
Im Glad to you liked it!
ۚۧÙŰȘÙÙÙÙ
Loved the way you explained. After a long time, I listened to some video for the whole duration.
That's wonderfull Hussein, Understanding the Why before going for a specific tech..
Thanks Rama!
The best content on the internet with a crazy presentation skills. it was wonderful Hussein.
Wow a whole ton of good stuff here. Props to you good sir!
DataSurgeon 369 đđ enjoy thanks for your comment
Thank you so much, Hussein! You're doing such a good job.
16:11 do you really want to do this with you life? I don't know man, I should have been a cook or something.
@Jamison Grotzinger I will kindly ask you to fuck off
Great video as usual. Great sense of humour and your explanations are very easy to follow.
Very good background knowledge leading into the sharding explanation. Nicely explained. Thank you!
Lol I love the way you explain stuff. Subscribed!
Awesome video Hussein , everytime i learn something new from your videos . Thanks for sharing your knowledge.
I am glad you are! Thanks for your comment
It's such a pleasure to watch! Fun and knowledgeable!
Another fantastic video from this great guy to start an another beautiful day!!!
Being a non-backend developer I still think atleast while watching your videos that I am a one of them !!! Superb and Easy Explaination
â€ïž thanks Ganesh!
I saw many videos but the way you explained horizontal/ vertical partitioning in just 30 sec is going to be with me foreever
Thanks a lot Hussein. Really nice and engaging video with awesome explaination
You deserve more views
Gonna share it with my peers.cheers
Aashish Goyal thank you đ
Great video! Thanks so much for taking the time to make this.
This channel is addictive
Fantastic video - thank you so much Hussein!
Great videos! Thank you for making them!
Thanks Tarik!
Just Wow! Learning so much from you Hussein
Great video. Love the way you explain the things. I wish you had elaborated little more or probably make a new video on why transactions are tough with sharding.
Correct that would require another video because its a deep topic , thanks for your comment â€ïž
knowledge we get here is pure gold. Thank you đ
When I find this DENSE content really enjoyable just like I'm watching Netflix, that does mean one thing you are really GREAT!!!, Thank you so much Hussein
I learned a lot by this. Thank you
Great video! I respect that you mentioned go
âPredictably Irrationalâ is a very good book! :-p
Hussein: You cannot longer perform transactions with sharding
Me: thanks captain, that's it for me on the subject
My man often mentions Django, I love it :)
Sharding is not really needed as you mentioned we can go for partitioning and local indexes within the partitions itself. What if writes are more there are various options in enterprise products, I worked long back with Oracle RAC setup with vplex managing the data storage for different nodes. Avoid writing component logging or audit trail logging to RDBMS instead write to nosql. I personally think only business data (OLTP) should be there in RDBMS all others can go log database (nosql) or splunk or datadog etc...
You are amazing . You are so wise. Thank u.
Glad to see you bro..I was wondering how you look like..!
loved it..thanks for the explanation
Good content for the basic understanding. Excellent
Glad it was helpful!
Great explanation
Your presentation is excellent
Excellent video. Thank you
Beautiful, thank you :)
Great content! Thank you! Do you have any particular resources for learning BE concepts? (except for a job :D)
Really great!
You are awesome. I have learned a lot from you. Thank you for being there.
But I have got a little confusion at 10:41. I agree the db requestes from clients will pass through the reverse proxy but the proxy should be SQL Aware like ProxySQL but can Nginx or HAProxy be configured as SQL Aware LB?
Excellent one... just wanted to know your approach to be on top of latest technology. Which resources do you use? Which technology podcast do you listen? Can you make video on this topic? đ
Awesome bro!
Excellent content bro
Sometimes you sound like Gru đ
Great content.
Can we use timescale DB extension with postgres for faster writes ?
My search ends here, to clear my doubt thank you.
One question I had does databases like Postgres provide facility for automatic sharding . If yes what process do they follow while doing the same
Great video Hussein. Can you please make a video on indexing and Acid transactions? That would be great! Thanks.
shaheer zaman thanks Shaheer! Check out my ACID video here Relational Database ACID Transactions (Explained by Example)
czcams.com/video/pomxJOFVcQs/video.html .. i am still need to work on the indexing video coming soon :)
thanks for the video
Cool thank you!
Life Saver!
@14:00 is it still the case? I read that mysql supports transaction on sharding xa transactions or distributed transactions. And mysql nda cluster support acid.
always great video Hussein, thank you. Journal, logging all the traffic for website/ application, could be Database heavy write , right?
Thanks Mohd! and Correct, logging is a database write heavy operation so you would choose an LSM based DB such as RocksDB or myRocks.. Check out my database engines video for more details on this topic
I'm your subscriber :D
The content is excellent for learning, really clears lots of stuff. But if anyone watching this video has an interview lined up(which is most probably true), don't explain in this fashion the strategy you choose to justify the problem you would be solving.
Hi
A very nice video.
I want to know the following:
I am have a "users" table with huge rows in it, it is having join with other 4 tables, what will happen If I partition the "Users" table. With the Joins and older queries will work as it is or I need to make changes in that, If yes what changes I need to do exactly to make the old queries work.
Thanks, Ill stick with sharting
I didn't get the difference between scaling writes by region vs sharding. Is it the same but the first one is sharding by region?
thank you
THOU SHALT NOT over-engineer too early. I am definitely putting that up on my wall :)
Why logic on client? You can make service that deals with this requests from client and send them to the right dB server? Or I don't understand something?
Just to be clear, you don't sacrifice the transactional capability per se; you can't do it across shards but if the sharding is designed to keep every shard with no client depedency on the other shards, the transactional capability isn't lost. You may want to think of it for any cloud company where the database sharding is done per tenant and every tenant has private data. Says shards are created on a range of tenant names (say a-c for shard1, d-e for shard2 etc)
, you aren't going to loose any capabilities here.
Amazing video, please make an arabic channel, we are missing this content here dude.
Please make a video on Elastic search
Just pointing this out: You can overwhelm a MySQL server's IO capacity very quickly with php-fpm at scale. Because each request spins up its own database connection, so unless you're throwing their write requests into a message queue then bulk inserting it's contents via some stateful service each connection = at least 1 IO event. And if you're co-locating odds are you don't want to set your io_capacity parameter above 2000 IOPS because of SSD burnout, and for php-fpm's concurrency model(or lack thereof) it's not hard to hit that threshold at even a small scale. Typically php-fpm sites of scale you'll see 30-50k IOPS, so even a cloud-based solution would be cost prohibitive.
Not trying to be a jerk or anything because this requires a lot of specific knowledge of a particular programming language's execution model. But having said that, it's totally possible to overwhelm IO without a few million daily page views. Since you gotta keep in mind 80% of the traffic hits in 20% of the day. That's also basically the mark when Web 2.0 companies of yesteryear started sharding their LAMP apps.
So just donât use fpm?
Good Content Hussein, but that will be Great if you use a whiteboard of some pictorial content instead of showing everything in Air.
Love your videos, what do you think of the 'New' SQL Databases such as CockroachDB and using them instead of complicating your life with sharding?
I need to do my research of them but I do this there are use cases for them. I still donât know what is new about them so canât really comment
@@hnasr Cool, I look forward to hearing your take on them, they're what I opt for in situations where I think I'd have to shard because sharding isn't fun. I've also tried others like Citus which are pretty great but a bit more work to get the most out of.
is there any risk of outdated data if the slaves do not sync on time? how do we deal with this?
Awesome video and thank you! What are 'file descriptors'? I think I heard you correctly at ~ 8:15.
Thank you!! File descriptors are handle to the TCP connection I donât know much about them (which is good means I need to read more about them and probably make a video)
Here is the wiki en.wikipedia.org/wiki/File_descriptor
Can you name the youtube's podcast you referred to?
Hello I have log server. There is not transaction. Its straight write and read. No update or delete will be performed. And no rollback is necessary. Problem is minimum log size is 400 GB plus which consists of raw text data only. DB size increases 40 GB per day min. So you understand there is a a lot of write in the database . Problem is there is too much write in one instance that I can not read the database. Its a MongoDB database and I searched and found I can do sharding to distribute the read write. And Mongo shard comes with tool like mongos which will distribute my query from client based on lets say timestamp. So should I proceed with this plan or should I do partition first? And also the pipeline is at development stage so if u recommend I can still change DB to postgres or other DB technology. Thanks.
Thats why i love mongo db, sharding Is a breeze but can be better with relations, i want it, i need ir. Mongo router rocks
is Replication increase reading capacity via Load balancing mechanism ???
Since I stackoverflow all months and cannot found any appropriate article about this. Even MongoDB Legacy Docs did not mention reading load balancing đą
Can we add a reverse proxy in between all our application servers and Sharded databases which handles all the routing logic? In this way, we can also avoid duplicate code in multiple application servers.
Yup! You are describing how Vitess was built and how many other apps before it. Good thinking đ
How about multi-master replication or bi-directional replication?
Just a quick question, how is 'scale writes by region' (using multiple master nodes in multiple regions and directing requests to nodes based on some key) different from sharding?
As he mentioned, While sharding ACID properties will not be applicable. Now in case of keeping data in regional database you still have ACID properties and you can sync the data later.
FYI, MongoDB have ACID transaction support with sharding
Great video! Are you planning to make a separated video about Vitess?
Nikola Kolarov thank you! Yes I am planning to make a video on Vitess
Hussein Nasser Awesome
So this is mostly focused on SQL right?
Do you have numbers that denote when a database is ready for partitioning, sharding etc?
there are no numbers i'm afraid. you have to know your system well enough to realize the problem and analyze what solution suits you best.
for example, if you are working with sensitive data, that you are legally bound to keep in certain locations, then you need sharding for that if you wanna expand to different regions.
one of the reasons numbers do not work is the nature of data itself. a table with 2 columns and 50 columns are going to have different query times. then it depends on the query itself. Also the technology you are working on etc.
so there are simply too many variables to have a standard.
I can tell you from experience, there was one system i was working on where we had around 200 mill rows in a single table. we didnt need any sort of partitions there.
On the other hand, i had a separate system, a little bit older tech as well, but data was vastly different. things came to a point where we needed partitions at 30 mill rows. index didnt cut it anymore.
the point is, it has to be determined for each system. we can have some guesses, but it wont be concrete.
I hope this makes sense :)
Greate content Hussen and thanks for sharing we learn a lot with you
I just comme back to a previous question asked by another person in comments, he sais : "just wanted to know your approach to be on top of latest technology. Which resources do you use? Which technology podcast do you listen? Can you make video on this topic? "
Can you answer to this please ? It will help us agains
Thanks again for this greate content
Momar Diouf thank you Momar! Appreciate you đ
I learn by listening to podcast, watching videos , reading and implementing the thing . I always ask why a tech exists before I ask what.
I made few videos on the topic of learning check them out
When Learning Backend Engineering Ask Why, not What (Minute Engineering)
czcams.com/video/67DglLwnBTU/video.html
My Preferred Method of Learning Backend Engineering Technologies
czcams.com/video/4NsWnT_-FoE/video.html
Learning at Home, Consistent Hashing, Empathy with Engineers and More - Software Chat
czcams.com/video/6PrR6SW4QGM/video.html
Advice for Junior backend engineers who just started their new jobs in software companies
czcams.com/video/V3C0VvNrFZ8/video.html
@@hnasr Thanks for the reply. I will check the links
Thanks
Your videos are great. Thank you
THANK you so much! appreciate it
Remember, when you are confused and you dont know why its slow, always start with distributed caching
When exactly though? Share some numbers
Bro you are fucking brilliant and that you are willing to teach others what you know in such a good engaging way I love it. God Bless you even if you don't believe in God or not. :)
What about the CQRS
sharding is like sharting, unpleasant but sometimes necessary
Just for my understanding, what is the difference between the videos presented here and the Udemy course? Is the Udemy course more code-oriented? Thanks
More detailed, structured with hands-on examples.
Hi Nasser,
For a system like Uber Eats , I believe we can split our Users, Orders, Restaurant, Dishes tables based on location. Now the question is shall we choose sql or nosql here? We generally go for noSQL because it avoid joins which can be costly in sharded DB, but if we maintain our tables on basis of location then there wont be any such case where we will join from table in another DB. So, which one shall we use???
Could you make a video about the 5G impact on (Cross-Shard Queries) knowing that 5G Latency is below 1 milliseconds..
1 - Could that bring sharding and cloud computing to the next level ?
2 - What will happen to data centers infrastructure, and also to database query languages in the future ?
3 - Will that reduce the data management cost for business owners ?
What amazing thoughts provoking questions Kapssul! Love them. That is going to take some time to research and answer because I have no idea. 5G is indeed a revolutionary tech and it will spin up the software engineering tech on its head.. thanks!!
@@hnasr I am on the way, you are already there.. you could grab the information quickly and more accurately, and share with us bit by bit your thoughts.. keep posting.. thanks for that quality content.. we learn a lot..
Itâs difficult to focus when the only thing I canât take my eyes off is the silver ps2
so my takeaway is
sharding is for scaling write queries (which is rare)
partition is for scaling read queries