PaulT

PFS pic

Better HDFS : #1 : PFS

PFS - Pragmatic POSIX compliant distributed filesystem



It's been battle tested (only Ubuntu so far, but should be *NIX agnostic)

  • PFS is similar to HDFS + POSIX layer
  • Logical /PFS directory is physically distributed between various hard drives
  • PFS allows (many) things that HDFS does not allow
  • PFS allows appends
  • PFS is not Java
  • PFS is implemented on a system level, so it can run in a public cloud
  • PFS could (in principle) run even on spot instances, but that would be silly
  • PFS backups/replicas are optional and can be done with simple rsync
  • If HW dies and leaves PFS in bad state (rare event) - it could be recovered with 'pfs fsck'
  • If 'pfs fsck' fails - PFS can be easily recovered manually
  • It is all possible, because PFS files are in fact just plain unix files
  • Currently I use /PFS to store (several months of) the fingerprints from my Cloud Servers, to do forensics if somebody has a success compromising my Cloud Servers
  • There is one controlling node 'aux' and any number of bricks
  • Not 100% of POSIX calls is currently implemented. Permissions are very relaxed, for example
  • It's all fuse, of course. I only had couple weeks between jobs - about a year ago
  • It is not "open source", but your super awesome company is welcome to give me a grant to take it open source, for example

Sample commands

root@n1:/PFS# ls /PFS
net  syslog  top
root@n1:/PFS# 

root@n1:/PFS# ls -l /PFS
total 12
drwxrwxrwx 2 root root 4096 Nov 20 00:00 net
drwxrwxrwx 2 root root 4096 Nov 20 00:00 syslog
drwxrwxrwx 2 root root 4096 Nov 20 00:00 top

root@n1:/PFS# ls -l `find /PFS` | head
-rw-rw-rw- 1 root root  402941 Mar 25  2014 /PFS/net/2014-03-25.node3
-rw-rw-rw- 1 root root  213784 Mar 25  2014 /PFS/net/2014-03-25.node4
-rw-rw-rw- 1 root root   82395 Mar 25  2014 /PFS/net/2014-03-25.node5
-rw-rw-rw- 1 root root 4827975 Mar 26  2014 /PFS/net/2014-03-26.node3
-rw-rw-rw- 1 root root 2559988 Mar 26  2014 /PFS/net/2014-03-26.node4
-rw-rw-rw- 1 root root 2566879 Mar 26  2014 /PFS/net/2014-03-26.node5
-rw-rw-rw- 1 root root 4817161 Mar 27  2014 /PFS/net/2014-03-27.node3
-rw-rw-rw- 1 root root 2586308 Mar 27  2014 /PFS/net/2014-03-27.node4
-rw-rw-rw- 1 root root 2713682 Mar 27  2014 /PFS/net/2014-03-27.node5
-rw-rw-rw- 1 root root 4806869 Mar 28  2014 /PFS/net/2014-03-28.node3

root@n1:/PFS# find /PFS | wc -l
1636
root@n1:/PFS# 

root@n1:/PFS# pfs fstatus
n2 -> 327
n3 -> 327
n4 -> 326
n5 -> 326
n6 -> 326
/PFS files: 1632
  DB total: 1632

root@n1:/PFS# pfs help
aux.create     
aux.start     
aux.stop     
create     
delete     
fs.mkdir $1    
fs.rmdir $1    
fsck $1    
fstatus     
help     
mount     
node $1 $2   
start     
status     
stop     
umount

November 19. 2014

Background

Update 

foreach *.csv -< F(3) | Z Rvi(60) | Lin(3) >- Avg
Only part of it can be expressed via current UNIX pipes. Note ( -< map and >- reduce ). The question of Hadoop was correct. The right answer - throw away HDFS! It's *not* part of the pipeline. M/R is.

The way I arrieved here was this. Today I started to write down some of the formulas, that are working in production. So I noticed the similarity with UNIX pipes, but there was no syntax to express some of the moves. First I thought <= for map and => for reduce, but wife said that -< and >- is more accurate. I agreed to that because for example => is already used in mathematics.

November 14. 2020

Figured out how to do version 2. blockchain is not bad for this kind of stuff actually.

March 6. 2021

Figured out how to replicate the same trick to entire Hadoop vertical. Not surprisingly. Scaling out is easy.

April 9. 2021

Figured out how to do it a bit better, by using one European open source component circa 2006. Also, there is now some blockchain stuff that pretends to be of use, only I'm undecided about it, because it looks like *everything* could have been expressed through that European component. Hadoop can definitely be expressed through two European components (one bashreduce analog, another part is FUSE part (circa 2006) + maybe Pfs, does not matter much) My pipes can not possibly be expressed (because need syntax to express semantics). Effectively, this page describes the next step after Hadoop. How much is Hadoop worth? Yes, this page costs millions. Only there is nobody left to read this page. Maybe in the future.

April 16. 2022

Whats better than Hadoop? Infinite Hadoop, of course.

May 12. 2022

Figured out how to make Pig bigger than Hadoop. The UNIX puzzle is a bit more interesting, though, so I am a bit undecided what to do with this Hadoop thing. Unix puzzle is here : Bayes. It's running since 1948 at least.

Dec 17. 2022

Figured out the (major) architectural problem in all the existing pragmatic solutions on this planet. Basically, memcached had the similiar problem only in reverse (overdone app layer) (still has it. nobody fixes stuff). Now need to think what to do about this all. The interesting moment is this. When there is a hard border to cross, because there is always "not enough time", smart person will always take the easier route. And that will of course backfire later, because it's all a skyscraper, so any shortcut in the foundation of the building will only show itself years later, and by that time it's impossible to fix the foundation. Any person, who builds the skyscrapers, is surely aware of that. Yet .. google Millennium Tower.

Mar 08. 2022

Figured out the perfect move. Had to blend several verticals together, but it *is* an FS. Ochestrated FS. The way they did Golden Gate bridge. I have everything to make this happen (and nobody else has, for some reason).

Jan 29. 2024

Dijkstra had the Self-Stabilization Ring. The last piece.

May 26. 2024