Location of some executables

Gordan Bobic gordan.bobic at gmail.com
Wed May 11 10:18:21 EDT 2011


On 05/11/2011 02:40 PM, briaeros007 wrote:

>>>> Apps are generally small. Data is generally big. I am all for putting
>>>> data
>>>> onto separate volumes, but when the entire OS install + apps is smaller
>>>> than
>>>> the amount of RAM on mid-range graphics card, I don't really see the gain
>>>> of
>>>> splitting it up in the general case.
>>>>
>>> Apps aren't small.
>>> If i install a Business Object, I must have  got 5 Go of free space to
>>> install it. Yes 5 Go.
>>> If i play with BMC remedy, it's the same idea. And data are in db.
>>
>> So separate that app's data store somewhere more sensible, not the entirety
>> of /usr.
> Hum...
> Data store are ALREADY in a db, so not with the apps (and on a separate server).

Are you saying that the entire 5GB of bloat is in the 
binaries/libraries? I don't (want to) believe that.

>>> Everybody doesn't use their system only as a desktop.
>>
>> Desktops can actually call for or at least justify _more_ partitioning than
>> servers, if you are using solid state storage.
>>
> sure...
> A desktop with ... 99% of the time only one session, no critical apps
> and onlye one  "nine" of availability (ie 90%) needs more care than
> critical servers.
> I don't agree, but it's only me after all ;)

This has nothing to do with availability - if you have lost your /usr, 
you've lost all your 9s of availability.

>>>>> - boot pxe for the system (one nfs share for the core system, and
>>>>> other for applications which are loaded when the system initialize
>>>>> itself).
>>>>> - etc...
>>>>
>>>> Sounds like an administrative nightmare. What commonly available package
>>>> management system will cope with that?
>>>>
>>> If you can't follow multiple nfs share, well pxe isn't for you.
>>> What commonly available package management system manage pxe ? none.
>>> I'm talking about specific need due to a boot procedure, and you talk
>>> about package management system. I don't see the links.
>>
>> So you are complaining that a completely custom brew setup you are running
>> doesn't agree with the default file locations in the ZFS package? And if you
>> are PXE booting with NFS root how does that relate to ZFS, exactly?
>>
>
> I don't complain.
> I never complained on this ml.
> You wanted use cases where /usr are not on the same partition. I give you one.

By citing NFS roots? Sorry, that doesn't wash. The situation is no 
different whether you are PXE booting a kernel off NFS or using local 
disks. The examples are analogous as far as separating /usr is concerned.

>>>>> /bin and /sbin are here for a reason : to provide a lightweigth
>>>>> environnement who can be used to manage the server, even if other
>>>>> partitions can't be mounted.
>>>>> I think It's a best practice to separate things on a server. if
>>>>> something bad happen, the other things aren't impacted.
>>>>
>>>> The chances are that you're not going to get very far in fixing things
>>>> without the things in /usr/bin and /usr/sbin.
>>>>
>>> So you're just saying that generations of sysadmin are just dumbass
>>> and protection/compartimentalization aren't useful ?
>>
>> I'm saying that times and resources have changed since this was a good idea.
>> Compartmentalization is a good thing if used appropriately. I do not believe
>> splitting every directory under / to a separate volume is appropriate for
>> any sane use-case I can think of.
>>
>> What was a good idea on SunOS 4.1.1 running off 40MB 9in SMD disks in 1990
>> isn't necessarily a good idea on the current generation of systems. Blindly
>> following pragmas deprecated by technology might imply some dumbassness
>> (your choice of words, not mine).
>
> I'd love to be a dumbass who apply choice wich simplify administration ;)
> And to have separate partition instead of a big / simplify the
> administration when you must act or diag  on system fs.
> (and you can do parallel fsck ^^)

You are _almost_ half way to having a useful feature of splitting things 
up with parallel fscking. But in case of ZFS we have no fsck (for better 
or worse), so in the case of the starting point of this argument, that's 
not a valid argument.

In all other cases, I still haven't heard an argument for why a 100MB 
root wouldn't be corrupted if the 5GB /usr is or vice versa.

>>> Let's just put all the things inside a big partition, and if something
>>> goes wrong, well, cry.
>>> Example : ->    the fs crash, and you lose all the data in the partition
>>> (and backup, as murphy law edicts, wasn't working this day)
>>> Did you prefer to lose just /usr, which are easily recreated, or /etc,
>>> /boot (can't boot withtout it ), /usr, /var and so on ?
>>
>> So by that logic, are you going to put /etc on a separate partition? Can you
>> list me a sensible use case where losing one wouldn't usually lead to losing
>> the lot? You can't argue that a solution to poor backups is partitioning
>> /usr off to a different partition (especially on the same backing storage,
>> be it disk, SAN, array, or whatever).
>>
> I love you.
> No sincerely.
> I give you just a simple example. And yet you use theses example as if
> it was all a new paradigm about system admin.
> Well, it was just what i say first : an example. perhaps not the best
> one, but only an example.

A _bogus_ example, dismantled as such, since it didn't support the point 
you were making.

> Poor backup IS a trouble that i've found in nearly ALL entities I've
> come across. And big entities!
> Multiple problem impacting one critical server at the same time exists.

I'm not disagreeing on either of those points. :)

> Hopefully, theses problemes represents less than 1% of normal platform
> but they do exits
> And when it strikes, you doesn't want to add more troubles to the incident.
 >
>> Since /usr isn't user-writable any more than / is, I don't really why losing
>> one is more likely than the other. And re-creating /usr without a backup is
>> going to mean a reinstall/rebuild/restore whatever you do. So IMO your
>> example is completely bogus.
>>
>
> /usr , and /etc are "apps writable" in some configuration.
> and /usr is more "every day" or application management, and "/" is
> more system management.
> The two things are not always on the same team.



>>> Other example : due to unforseen evenement, you must add 4 Go to /usr or
>>> /var.
>>> did you want to take the risk to do an "resize2fs" on the root
>>> partition and an modification of physical partition, or just lvextend
>>> and resize2fs on a non root partition?
>>
>> Why would you possibly be rationing space that tightly? Disk space is worth
>> £0.04/GB. So be extravagant - splash out an extra £4 in disk space at
>> installation time and have a 100GB /. More than you'll even sanely use up
>> for an app installation that you hadn't anticipated at the time when you
>> built the system.
>>
> Excuse me ?
> When you ask a server you are always working
> - without recommandation or technical procedure
> - can command any hardware which exists
> - add any options to it
> or you are restrained in your choice of hardware, software AND configuration ?

Usually the constraints are not such that they cause bad design 
decisions. Most tasks fall into two categories:

1) Keeping existing systems running

This usually doesn't involve dumping gigabytes of apps on an existing 
system in production.

2) Deploying and designing new systems

If that is what I am doing, then yes, I get to spec the hardware and 
write the procedures.

What you are suggesting is akin to saying that it is OK to try to fix 
every problem with a hammer because it's the only tool you have been given.

> (and SAS disk aren't 0.04/GB but budget isn't really the point here.
> When i've got a blade with 70 Go disk to mount an oracle server with
> 50 GB of datafile ... I wouldn't create a 100 GB /usr. And experience
> learns me that you must always let some unallocated free space with
> lvm for subsequent demands)

Personally, I think LVM is _evil_ when it comes to any kind of 
performance optimization. It destroys in one fell swoop any careful disk 
layout optimization you may have carried out if you are on any kind of a 
RAID setup. If you carefully pick your chunk size, stripe-width and 
block-group size, you can get a lot of extra performance out of your 
disks (if you get it wrong you can reduce the random file read 
performance of your array down to the performance of a single disk). LVM 
headers upset this, but that is manageable. What isn't manageable is 
when you start using LVM to stretch a volume to another disk since you 
can no longer know how everything lines up underneath (note RAID 
reshaping will throw things out in the same way).

> Actually, the technical procedure i've got to use rhel is
> / : 1G
> /home : 1G
> /opt : 1G
> /tmp : 1G
> /var : 2G
> /usr : 2G
>
> I don't decide if I want to make more or less. My enterprise says "you
> create the system with theses settings, and only after you adapt".

Right - so we have gone from "I think this is the best way to do things" 
to "this is the way I have to do things due to a dictat". The latter is 
a political issue I'm not going to discuss. It is the former that I was 
arguing against.

But with the numbers you list above, you'd still be in trouble with your 
4GB app on /usr, even though it is separate. So the point is moot.

>>> In all corporations I work, there was never a "big /". And this for
>>> linux, aix, solaris or hp-ux.
>>
>> My experience is varied - very limited purpose hardware (server farms) small
>> root is relatively common (then again, in a thin majority of such setups
>> I've worked on Solaris is in use, and as mentioned previously, that has /usr
>> implicitly on /).
>>
>> But I have _never_ seen a call to suddenly and unexpectedly dump an extra
>> 4GB of _apps_ on an existing system. That's just poor planning. And
>> regardless, if you are making a change of that magnitude, you'll typically
>> jumpstart/kickstart a new server, test the new app, the migrate to live, not
>> just randomly dump gigabytes of apps onto a system that is sufficiently
>> narrow-purposed to have a tiny root fs.
>>
>
> I have already seen this one.
>
> You've got a server which was correctly adapted for an apps.
> you install the apps, you do all the configuration with specific
> procedure and so on.
> Well pretty though stuff.
> All works ok.
> You manage the server normally, and follow what the client ask. And
> they asked to add this apps, this account and so on the webserver.
 >
> And one day you must add a security/... update of the apps (or another
> apps, or ..).
> the update ask you 5 Go of free space
> But you don't have 5 Go. What do you do ?
> 1°) you recreate a new server and lost one weeks of works ?

If it takes you a week to build a (replacement) server, you're doing it 
wrong.

And you haven't actually mentioned the testing of such massive changes 
on a production system. You wouldn't expect them to "just work", would you?

You can't argue constraints of enterprise level setups without having 
enterprise level procedures.

> 2°) you say to the client :  "i can't do this, technically i could but
> i doesn't want since it's not up my standards"
> 3°) you say to the client  "well, it's not optimal, but i could do
> this and that to make it work. Are you ok with that?"

Most of my work comes from clients who have to deal with fallout and 
consequence of somebody having done option 3.

>>> And the fact that isn't on "text install but on gui" doesn't seem
>>> relevant to me.
>>> Any server now can use a gui, and it's been some times we don't do
>>> console install (or in specific hardware).
>>
>> So what do ancient text-only installers you mentioned have to do with
>> anything, then?
>>
> I'm afraid my explanation wasn't clear.
> If I used  "true (plain old?) installer" was to show two things
> 1°) that it's a policy which was well established and wasn't just some
> folks fancy.

Well established based on what was useful and appropriate 20 years ago. 
Not necessarily based on what is useful and appropriate today.

> 2°) to not take into account desktop centric installer.

RHEL6 (what I mentioned in the example) isn't a desktop centric distro.

And from what you said about your distro - there being no default distro 
kernel - well, that implies a distro that isn't even a desktop one, let 
alone a server one.

Gordan



More information about the zfs-discuss mailing list