fsync()-in webdav PUT

[Maxim Dounin]

Hello!

On Fri, Mar 02, 2018 at 10:12:02AM +0100, Nagy, Attila wrote:

> On 02/28/2018 03:08 PM, Maxim Dounin wrote:
> > The question here is - why you want the file to be on disk, and
> > not just in a buffer?  Because you expect the server to die in a
> > few seconds without flushing the file to disk?  How probable it
> > is, compared to the probability of the disk to die?  A more
> > reliable server can make this probability negligible, hence the
> > suggestion.
> Because the files I upload to nginx servers are important to me. Please 
> step back a little and forget that we are talking about nginx or an HTTP 
> server.

If file are indeed important to you, you have to keep a second 
copy in a different location, or even in multiple different 
locations.  Trying to do fsync() won't save your data in a lot of 
quite realistic scenarios, but certainly will imply performance 
(and complexity, from nginx code point of view) costs.

> We have data which we want to write to somewhere.
> Check any of the database servers. Would you accept a DB server which 
> can loose confirmed data or couldn't be configured that way that a 
> write/insert/update/commit/whatever you use to modify or put data into 
> it operation is reliably written by the time you receive acknowledgement?

The "can loose confirmed data" claim applies to all database 
servers, all installations in the world.  There is no such thing 
as 100% reliability.  And the question is how probable data loss 
is, and if we can ignore a particular probability or not.

> Now try to use this example. I would like to use nginx to store files. 
> That's what HTTP PUT is for.
> Of course I'm not expecting that the server will die every day. But when 
> that happens, I want to make sure that the confirmed data is there.
> Let's take a look at various object storage systems, like ceph. Would 
> you accept a confirmed write to be lost there? They make a great deal of 
> work to make that impossible.
> Now try to imagine that somebody doesn't need the complexity of -for 
> example- ceph, but wants to store data with plain HTTP. And you got 
> there. If you store data, then you want to make sure the data is there.
> If you don't, why do you store it anyways?

So, given the fact that there is no such thing as 100% 
reliability, you suggest to do not store files at all?  I don't 
think it's viable approach - and clearly you are already doing the 
opposite.  Rather, you want to consider various scenarios and 
their probabilities, and minimize probabilities of loosing data 
where possible and makes sense.

And that's why I asked you to compare the probability you are 
trying to avoid with other probabilities which can cause data loss - 
for example, the probability of the disk to die.  Just for the 
reference, assuming you are using a commodity HDD to store your 
files, the probability that it will fail within a year is about 2% 
(see, for example, Backblaze data, recent stats are available at
https://www.backblaze.com/blog/hard-drive-stats-for-2017/).

Moreover, even if you have numbers on hand and this numbers will 
show that you indeed need to ensure syncing files to disk to reach 
greater reliability, doing fsync() might not be the best way to 
achieve this.  For example, doing sync() instead after loading 
multiple files might be a better solution, both due to lower 
complexity and higher performance.

-- 
Maxim Dounin
http://mdounin.ru/