Your SSD lies but that's ok .. I think

Your SSD lies but that's ok .. I think | Postgres fsync

Рет қаралды 19,816

Күн бұрын

fsync is a linux system call that flushes all pages and metadata for a given file to the disk. It is indeed an expensive operation but required for durability especially for database systems. Regular writes that make it to the disk controller are often placed in the SSD local cache to accumulate more writes before getting flushed to the NAND cells.
However when the disk controller receives this flush command it is required to immediately persist all of the data to the NAND cells.
Some SSDs however don't do that because they don't trust the host and no-op the fsync, in some cases SSD buffers it for few microseconds to receive more writes and hope it doesn't crash in this microsecond. Others battery powered SSDs can completely ignore host flushes because it can survive a crash. In this video I explain this in details and go through details on how postgres provide so many options to fine tune fsync
0:00 Intro
1:00 A Write doesn’t write
2:00 File System Page Cache
6:00 Fsync
7:30 SSD Cache
9:20 SSD ignores the flush
9:30 15 Year old Firefox fsync bug
12:30 What happens if SSD loses power
15:00 What options does Postgres exposes?
15:30 open_sync (O_SYNC)
16:15 open_datasync (O_DSYNC)
17:10 O_DIRECT
19:00 fsync
20:50 fdatasync
21:13 fsync = off
23:30 Don’t make your API simple
26:00 Database on metal?
Fundamentals of Backend Engineering Design patterns udemy course (link redirects to udemy with coupon)
backend.hussei...
Fundamentals of Networking for Effective Backends udemy course (link redirects to udemy with coupon)
network.hussei...
Fundamentals of Database Engineering udemy course (link redirects to udemy with coupon)
database.husse...
Follow me on Medium
/ membership
Introduction to NGINX (link redirects to udemy with coupon)
nginx.husseinn...
Python on the Backend (link redirects to udemy with coupon)
python.hussein...
Become a Member on KZfaq
/ @hnasr
Buy me a coffee if you liked this
www.buymeacoff...
Arabic Software Engineering Channel
/ @husseinnasser
🔥 Members Only Content
• Members-only videos
🏭 Backend Engineering Videos in Order
backend.hussei...
💾 Database Engineering Videos
• Database Engineering
🎙️Listen to the Backend Engineering Podcast
husseinnasser....
Gears and tools used on the Channel (affiliates)
🖼️ Slides and Thumbnail Design
Canva
partner.canva....
Stay Awesome,
Hussein

Пікірлер: 34

@hnasr Жыл бұрын

get my database course database.husseinnasser.com

@mossaabboudchicha84 Жыл бұрын

Hello hussein,very good explanation,i juste want to advise to mention the case of fsync failure that has been discovered by postgres developers during avril 2018 where they have been based on incorrect assumption of the behaviour of the fsync kernel system call,here is the link of the talk : kzfaq.info/get/bejne/Z7yHfKpztrLYqKM.html

@WinterHoax Жыл бұрын

How big is the course?

@michaelutech4786 Жыл бұрын

"Let the user feel it" - man, I love you. Can't remember when I enjoyed a tech video as much as this wonderful ranting session! I'm working in this job for more than 30 years now. My mantra became "don't throw away information" and "find the right abstraction level". You don't use these expressions, but it looks like you needed much less than two decades to come to a similar conclusion.

@durjaarai7737 Жыл бұрын

Its an unknown area that most engineers tend to look other way. Thanks for doing a deep dive. I love your content. More power to you!!

@saeedalobidi4195 Жыл бұрын

"you have to show everything in abstract layer" i can't agree more

@hakanaki 8 ай бұрын

Hi Hussein, you should create a podcast. I love listening to your tech talk while working.

@michaelutech4786 Жыл бұрын

It's funny to listen to these thoughts. "A database does not really need a file system" or "a database as a file system". When I first worked with databases, it was good practice to assign a database a disk partition for performance and reliability. I also remember that I think Microsoft played with the idea to integrate SQL server as filesystem in windows. Some dinosaur OS actually had filesystems that used database concepts. When I first saw Emacs, I had the impression that this is actually an OS integrated in my editor. Chrome is an operating system with a UI integrated in a Browser. We are using hypertext transfer protocol as a network layer and remote procedure communication layer. The problem really seems to be that we just can't agree on the necessity to agree on something. And then we make technical choices based on politics.

@blehbleh9283 Жыл бұрын

3:20 A DB as an operating system would be interesting but you'd be probably reimplementing a lot of the same work with disk drivers that the kernel does

@blehbleh9283 Жыл бұрын

28:16 well I guess not if you design for specific hardware

@yapet Жыл бұрын

“There is no need for databases to use a filesystem”. Welp, technically correct is the best kind of correct. Filesystem is an abstraction layer between disk drivers and you issuing `write` syscalls. If not for that, database would have to implement drivers for every protocol it might be used with (and it will be used with a LOT of them). SATA, PCI-E, SAS, NVME, USB(?), SCSSI (okay, this is a stretch), DVD-RAM (what point am I even trying to make?), ZFS, network mounted drives, SMB, FTP, some crazy drive-over-ethernet solution, etc. Maybe filesystem is not a efficient enough abstraction for databases, maybe we need another one. Or maybe not. Maybe heavily vertically integrated cloud service providers, where they know exactly which hardware, disk topology, etc. they will be running on, could specialize to that hardware. Maybe implementing the most used protocols, like SATA, SAS, NVME, and falling back onto the os fs if not supported would make sense, but it is a monumental task to implement, at least as good as the os drivers. And there also all kinds of RAID solutions. There are things like ZFS evolving on its own, at a rapid pace. Do we want our database to break, when we update ZFS, and open source maintainers haven’t gotten to patch it yet? Not sure possible elimination of os fs inefficiencies is worth it. Although, I am a complete dum-dum when it comes to the file system implementations. Maybe there is something there. Anyway, this might be SO factually incorrect (once again, I’m a dum-dum) but these are the initial concerns I have right now, without researching the topic any further.

@MaulikParmar210 Жыл бұрын

It's not syscalls that's inefficient, it's the tradeoffs that DB engine makes while choosing to load store data from location via single file / multi file / seeks / pages that make it inefficient and very very efficient in many ways. Filesystem is juat in interface similar to TCP / IP on top of OSI network stack implementation in an operating system. Overhead on drivers is minimal i.e. they work in order to allow access hardware in a systematic manner, withiyt drivers it wouldn't be possible to interface at all. Filesystem API consists of kernel abstractions that would map syscalls to driver calls and then driver code will work efficiently to read / write location. It doesn't know things about fs tree / journel or file metadata and have the ability to make decisions on its own. It just knows block and operarion at the most. DB engines can choose to issue operarions based on disk type to optimize as each hardware physically behaves differently. But as usual, DB engines usually choose to go generic to avoid tradeoffs that come with working with lower level access patterns. OS / Drviers code is simply an agent that translates resource calls without flooding hardware or corrupting data. Your program can choose to take advantage of patterned access or dumb access. P.S. when we talk about large scale ETL you can actually see filesystem becomes integral part of tooling rather than just means of storage mechanism, there are solutions on top of FS that computes data - map reduces it and returms back data to avoid shifting tons of stored data. Traditional DBs will never make use of disk based optimizations in general, atleast not OSS ones. Enterprise solutions already make use of such optimizations to squeeze out performance. DMA requires external controller to access memory, it's mainly used for IO intrrupts i.e. mouse move calls without CPU context switching to write that data and then handle intrrupt - DBs can't use DMA at hardware level as there's no hardware that would trigger DMA - disks can choose to use any UDMA modes but that's upto driver and hardware not between OS and Application which is independent of these impmentation details.

@btom1990 Жыл бұрын

Huh? At least in Linux there is still an abstraction layer between the filesystem and having to talk to HDDs, SSDs, RAID controllers etc. By using a block device you can still have RAID and don't need to deal with filesystem semantics. That's exactly what CEPHs BlueStore backend is doing.

@MaulikParmar210 Жыл бұрын

@Thomas Butz BlueFS is still an abstraction on top of syscalls. Different filesystem impmentation on top of basic read / write calls, that every other ( XFS, ZFS, EXT, NTFS, NFS ) does. Nothing new, just keeping journel in separate DB outside of block storage instead of on device. It's still the same as using traditional filesystem, except ceph now knows early, which page / block to read without managing multiple journels and reading / writing back and forth all the time. Just have a look at the source in the block device, which is based on all other types of fs. It still uses posix based system calls to manage actual files / data. Whatever talked in this video is very specialised HW impmentation where OS and almost all management stack is removed from software just to run single specific software, in that case many scalable storage is available, the problem is cost at scale, you can easily get away with commoditiy HW, ceph is designed to run on comodity HW mainly xfs as underlying filesystem, but can also support standard syscalls via vfs abstraction. It still uses the same abstraction that's talked about here! P.S. check ceph/src/os/ in repository RAID is HW level abstraction - Bluestore on top of xfs doesn't know what's underlying vfs abstraction. That's the whole point of having OS managing RAID in the driver codebase at kernel level along with bios to talk to the device at the given bus address.

@christianonyango7460 Жыл бұрын

every morning I have to watch one of your videos before I start my day :)

@kap1840 Жыл бұрын

There are multiple DBs that bypass the file systems.

@blehbleh9283 Жыл бұрын

The more you learn low level OS kernel networking stack stuff, the more you see algorithms (that are fundamnetally the same) going by different names as well

@techdemy4050 Жыл бұрын

Better title of the video can be "Love the way your SSDs Lie" lol

@WinterHoax Жыл бұрын

Please make a video about LSM tree and the compaction strategies of Cassandra and how the compaction works with less than 50% disk overhead

@bephrem Жыл бұрын

this is so fire

@kuhluhOG Жыл бұрын

3:30 I actually heard once that some companies experiment with that idea. They still run on top of an OS (it would be insane to implement essentially their own OS ofc), but have "direct" access to the device (well, as direct as it is possible).

@tharunthennarasu2839 10 ай бұрын

isnt xbox did this

@electronlabs2802 Жыл бұрын

If you cut the os and started to talk to hardware directly you might do something shaddy and disturb the file system i guess. I dont even know how it would work if the operation needs some syncing with the OS commands. It would be great to know if such thing is possible ;)

@MyEconomics101 Жыл бұрын

Time for Google & Co to make their own SSD for the data centre? Would not be surprised.

@MartinPHellwig Жыл бұрын

Look up Oracle and raw partition, I remember needing to do that a couple of decades ago.

@Zuriki09 11 ай бұрын

Actually this is a common problem with large scale NAS/SAN systems. SSDs lie about writing data, power goes out, data is lost. You need specific types of (expensive) SSD that properly flush data on power loss.

@lepidoptera9337 9 ай бұрын

That's not how you prevent data loss. You have a log that has to be updated after a successful file IO operation. If the operation is not marked as finished in the log, you can't rely on it. Journaling file systems have been doing that for you for ages. Not sure why anybody thinks they have to re-invent a half century old (or older) technology here.

@Zuriki09 9 ай бұрын

@@lepidoptera9337 the problem is SSDs lie about the write operation being finished and don't have capacitors that can allow for finishing writes on power loss.

@LewisCampbellTech 11 ай бұрын

I'm pretty sure tigerbeetleDB bypass the filesystem entirely and do direct IO.

@kumailn7662 Жыл бұрын

It would be better if you use diagrams to explain instead of moving your hand and fingers, the way you blend and drag the words, it would be better if you speak clearly. I like your topics and explanation and I like to learn from you but this annoying blended language really terrible some time to understand.

@Shwed1982 Жыл бұрын

Imaging the slides. Like a box and label SSD on top of it to watch for 5 minutes. Not an excellent idea 😂