How to safely maintain a status file

[Richard Baron Penman]

Hello,

I want my script to generate a ~1KB status file several times a second.
The script may be terminated at any time but the status file must not
be corrupted.
When the script is started next time the status file will be read to
check what needs to be done.

My initial solution was a thread that writes status to a tmp file
first and then renames:

open(tmp_file, 'w').write(status)
os.rename(tmp_file, status_file)

This works well on Linux but Windows raises an error when status_file
already exists.
http://docs.python.org/library/os.html#os.rename


I guess I could delete the status file:

open(tmp_file, 'w').write(status)
if os.path.exists(status_file):
os.remove(status_file)
os.rename(tmp_file, status_file)

and then on startup read from tmp_file if status_file does not exist.
But this seems awkward.


Is there a better way? Or do I need to use a database?

Richard

[Christian Heimes]

Am 08.07.2012 13:29, schrieb Richard Baron Penman:
> My initial solution was a thread that writes status to a tmp file
> first and then renames:
>
> open(tmp_file, 'w').write(status)
> os.rename(tmp_file, status_file)

You algorithm may not write and flush all data to disk. You need to do
additional work. You must also store the tmpfile on the same partition
(better: same directory) as the status file

with open(tmp_file, "w") as f:
f.write(status)
# flush buffer and write data/metadata to disk
f.flush()
os.fsync(f.fileno())

# now rename the file
os.rename(tmp_file, status_file)

# finally flush metadata of directory to disk
dirfd = os.open(os.path.dirname(status_file), os.O_RDONLY)
try:
os.fsync(dirfd)
finally:
os.close(dirfd)

> This works well on Linux but Windows raises an error when status_file
> already exists.
> http://docs.python.org/library/os.html#os.rename

Windows doesn't suppport atomic renames if the right side exists. I
suggest that you implement two code paths:

if os.name == "posix":
rename = os.rename
else:
def rename(a, b):
try:
os.rename(a, b)
except OSError, e:
if e.errno != 183:
raise
os.unlink(b)
os.rename(a, b)

Christian

[Michael Hrivnak]

What are you keeping in this status file that needs to be saved
several times per second? Depending on what type of state you're
storing and how persistent it needs to be, there may be a better way
to store it.

Michael

> --
> http://mail.python.org/mailman/listinfo/python-list

[Laszlo Nagy]

On Sun, 8 Jul 2012 21:29:41 +1000, Richard Baron Penman
<rich...@gmail.com> declaimed the following in gmane.comp.python.general:

>> and then on startup read from tmp_file if status_file does not exist.
>> But this seems awkward.
>>

> It also violates your requirement -- since the "crash" could take
> place with a partial "temp file".
>
> I'd suggest that, rather than deleting the old status file, you
> rename IT -- and only delete it IF you successfully rename the temp
> file.
Yes, this is much better. Almost perfect. Don't forget to consult your
system documentation, and check if the rename operation is atomic or
not. (Most probably it will only be atomic if the original and the
renamed file are on the same physical partition and/or mount point).

But even if the rename operation is atomic, there is still a race
condition. Your program can be terminated after the original status file
has been deleted, and before the temp file was renamed. In this case,
you will be missing the status file (although your program already did
something just it could not write out the new status).

Here is an algorithm that can always write and read a status (but it
might not be the latest one). You can keep the last two status files.

Writer:
* create temp file, write new status info
* create lock file if needed
* flock it
* try:
* delete older status file
* rename temp file to new status file
* finally: unlock the lock file

Reader:

* flock the lock file
* try:
* select the newer status file
* read status info
* finally: unlock the lock file

It is guaranteed that you will always have a status to read, and in most
cases this will be the last one (because the writer only locks for a
short time). However, it is still questionable, because your writer may
be waiting for the reader to unlock, so the new status info may not be
written immediatelly.

It would really help if you could tell us what are you trying to do that
needs status.

Best,

Laszlo

[Christian Heimes]

Am 08.07.2012 22:57, schrieb Laszlo Nagy:
> But even if the rename operation is atomic, there is still a race
> condition. Your program can be terminated after the original status file
> has been deleted, and before the temp file was renamed. In this case,
> you will be missing the status file (although your program already did
> something just it could not write out the new status).

You are contradicting yourself. Either the OS is providing a fully
atomic rename or it doesn't. All POSIX compatible OS provide an atomic
rename functionality that renames the file atomically or fails without
loosing the target side. On POSIX OS it doesn't matter if the target exists.

You don't need locks or any other fancy stuff. You just need to make
sure that you flush the data and metadata correctly to the disk and
force a re-write of the directory inode, too. It's a standard pattern on
POSIX platforms and well documented in e.g. the maildir RFC.

You can use the same pattern on Windows but it doesn't work as good and
doesn't guaranteed file integrity for two reasons:

1) Windows's rename isn't atomic if the right side exists.

2) Windows locks file when a program opens a file. Other programs can't
rename or overwrite the file. (You can get around the issue with some
extra work, though.)

Christian

[Plumo]

> What are you keeping in this status file that needs to be saved
> several times per second?  Depending on what type of state you're
> storing and how persistent it needs to be, there may be a better way
> to store it.
>
> Michael

This is for a threaded web crawler. I want to cache what URL's are
currently in the queue so if terminated the crawler can continue next
time from the same point.

[Plumo]

> > and then on startup read from tmp_file if status_file does not exist.
> > But this seems awkward.
>
>         It also violates your requirement -- since the "crash" could take
> place with a partial "temp file".

Can you explain why?
My thinking was if crash took place when writing the temp file this
would not matter because the status file would still exist and be read
from. The temp file would only be renamed when fully written.

[Plumo]

> Windows doesn't suppport atomic renames if the right side exists.  I
> suggest that you implement two code paths:
>
> if os.name == "posix":
>     rename = os.rename
> else:
>     def rename(a, b):
>         try:
>             os.rename(a, b)
>         except OSError, e:
>             if e.errno != 183:
>                 raise
>             os.unlink(b)
>             os.rename(a, b)

Problem is if the process is stopped between unlink and rename there
would no status file.

[Christian Heimes]

Am 09.07.2012 07:50, schrieb Plumo:
>> Windows doesn't suppport atomic renames if the right side exists. I
>> suggest that you implement two code paths:
>

> Problem is if the process is stopped between unlink and rename there
> would no status file.

Yeah, you have to suffer all of Windows' design flaws. You could add a
backup status file or use a completely different approach.

[Nobody]

On Sun, 08 Jul 2012 22:57:56 +0200, Laszlo Nagy wrote:

> Yes, this is much better. Almost perfect. Don't forget to consult your
> system documentation, and check if the rename operation is atomic or not.
> (Most probably it will only be atomic if the original and the renamed file
> are on the same physical partition and/or mount point).

On Unix, rename() is always atomic, and requires that source and
destination are on the same partition (if you want to "move" a file across
partitions, you have to copy it then delete the original).

> But even if the rename operation is atomic, there is still a race
> condition. Your program can be terminated after the original status file
> has been deleted, and before the temp file was renamed. In this case, you
> will be missing the status file (although your program already did
> something just it could not write out the new status).

In the event of abnormal termination, losing some data is to be expected.
The idea is to only lose the most recent data while keeping the old copy,
rather than losing everything. Writing to a temp file then rename()ing
achieves that.

[Duncan Booth]

Richard Baron Penman <rich...@gmail.com> wrote:

> Is there a better way? Or do I need to use a database?

Using a database would seem to meet a lot of your needs. Don't forget that
Python comes with a sqlite database engine included, so it shouldn't take
you more than a few lines of code to open the database once and then write
out your status every few seconds.

import sqlite3

con = sqlite3.connect('status.db')

...
with con:
cur = con.cursor()
cur.execute('UPDATE ...', ...)

and similar code to restore the status or create required tables on
startup.

--
Duncan Booth http://kupuguy.blogspot.com

[John Nagle]

On 7/8/2012 2:52 PM, Christian Heimes wrote:
> You are contradicting yourself. Either the OS is providing a fully
> atomic rename or it doesn't. All POSIX compatible OS provide an atomic
> rename functionality that renames the file atomically or fails without
> loosing the target side. On POSIX OS it doesn't matter if the target exists.

Rename on some file system types (particularly NFS) may not be atomic.

>
> You don't need locks or any other fancy stuff. You just need to make
> sure that you flush the data and metadata correctly to the disk and
> force a re-write of the directory inode, too. It's a standard pattern on
> POSIX platforms and well documented in e.g. the maildir RFC.
>

> You can use the same pattern on Windows but it doesn't work as good.

That's because you're using the wrong approach. See how to use
ReplaceFile under Win32:

http://msdn.microsoft.com/en-us/library/aa365512%28VS.85%29.aspx

Renaming files is the wrong way to synchronize a
crawler. Use a database that has ACID properties, such as
SQLite. Far fewer I/O operations are required for small updates.
It's not the 1980s any more.

I use a MySQL database to synchronize multiple processes
which crawl web sites. The tables of past activity are InnoDB
tables, which support transactions. The table of what's going
on right now is a MEMORY table. If the database crashes, the
past activity is recovered cleanly, the MEMORY table comes back
empty, and all the crawler processes lose their database
connections, abort, and are restarted. This allows multiple
servers to coordinate through one database.

John Nagle

[Michael Hrivnak]

Please consider batching this data and doing larger writes. Thrashing
the hard drive is not a good plan for performance or hardware
longevity. For example, crawl an entire FQDN and then write out the
results in one operation. If your job fails in the middle and you
have to start that FQDN over, no big deal. If that's too big of a
chunk for your purposes, perhaps break each FQDN up into top-level
directories and crawl each of those in one operation before writing to
disk.

There are existing solutions for managing job queues, so you can
choose what you like. If you're unfamiliar, maybe start by looking at
celery.

Michael

> --
> http://mail.python.org/mailman/listinfo/python-list

[Christian Heimes]

Am 09.07.2012 22:24, schrieb John Nagle:
> Rename on some file system types (particularly NFS) may not be atomic.

The actual operation is always atomic but the NFS server may not notify
you about success or failure atomically.

See http://linux.die.net/man/2/rename, section BUGS.

> That's because you're using the wrong approach. See how to use
> ReplaceFile under Win32:
>
> http://msdn.microsoft.com/en-us/library/aa365512%28VS.85%29.aspx

The page doesn't say that ReplaceFile is an atomic op.

Christian

[alex23]

On Jul 10, 6:24 am, John Nagle <na...@animats.com> wrote:
> That's because you're using the wrong approach. See how to use
> ReplaceFile under Win32:
>
> http://msdn.microsoft.com/en-us/library/aa365512%28VS.85%29.aspx

I'm not convinced ReplaceFile is atomic:

"The ReplaceFile function combines several steps within a single
function. An application can call ReplaceFile instead of calling
separate functions to save the data to a new file, rename the original
file using a temporary name, rename the new file to have the same name
as the original file, and delete the original file."

About the best you can get in Windows, I think, is MoveFileTransacted,
but you need to be running Vista or later:

http://msdn.microsoft.com/en-us/library/windows/desktop/aa365241(v=vs.85).aspx

I agree with your suggestion of using something transactional that
isn't bound to later Window versions, though.

[Laszlo Nagy]

> You are contradicting yourself. Either the OS is providing a fully
> atomic rename or it doesn't. All POSIX compatible OS provide an atomic
> rename functionality that renames the file atomically or fails without
> loosing the target side. On POSIX OS it doesn't matter if the target exists.

This is not a contradiction. Although the rename operation is atomic,
the whole "change status" process is not. It is because there are two
operations: #1 delete old status file and #2. rename the new status
file. And because there are two operations, there is still a race
condition. I see no contradiction here.

>
> You don't need locks or any other fancy stuff. You just need to make
> sure that you flush the data and metadata correctly to the disk and
> force a re-write of the directory inode, too. It's a standard pattern on
> POSIX platforms and well documented in e.g. the maildir RFC.

It is not entirely true. We are talking about two processes. One is
reading a file, another one is writting it. They can run at the same
time, so flushing disk cache forcedly won't help.

[Laszlo Nagy]

> Renaming files is the wrong way to synchronize a
> crawler. Use a database that has ACID properties, such as
> SQLite. Far fewer I/O operations are required for small updates.
> It's not the 1980s any more.

I agree with this approach. However, the OP specifically asked about
"how to update status file".

[Christian Heimes]

Am 12.07.2012 14:30, schrieb Laszlo Nagy:
> This is not a contradiction. Although the rename operation is atomic,
> the whole "change status" process is not. It is because there are two
> operations: #1 delete old status file and #2. rename the new status
> file. And because there are two operations, there is still a race
> condition. I see no contradiction here.

Sorry, but you are wrong. It's just one operation that boils down to
"point name to a different inode". After the rename op the file name
either points to a different inode or still to the old name in case of
an error. The OS guarantees that all processes either see the first or
second state (in other words: atomic).

POSIX has no operation that actually deletes a file. It just has an
unlink() syscall that removes an associated name from an inode. As soon
as an inode has no names and is not references by a file descriptor, the
file content and inode is removed by the operating system. rename() is
more like a link() followed by an unlink() wrapped in a system wide
global lock.

> It is not entirely true. We are talking about two processes. One is
> reading a file, another one is writting it. They can run at the same
> time, so flushing disk cache forcedly won't help.

You need to flush the data to disk as well as the metadata of the file
and its directory in order to survive a system crash. The close()
syscall already makes sure that all data is flushed into the IO layer of
the operating system.

With POSIX semantics the reading process will either see the full
content before the rename op or the full content after the rename op.
The writing process can replace the name (rename op) while the reading
process reads the status file because its file descriptor still points
to the old status file.

Christian

[Hans Mulder]

On 12/07/12 14:30:41, Laszlo Nagy wrote:
>> You are contradicting yourself. Either the OS is providing a fully
>> atomic rename or it doesn't. All POSIX compatible OS provide an atomic
>> rename functionality that renames the file atomically or fails without
>> loosing the target side. On POSIX OS it doesn't matter if the target
>> exists.

> This is not a contradiction. Although the rename operation is atomic,
> the whole "change status" process is not. It is because there are two
> operations: #1 delete old status file and #2. rename the new status
> file. And because there are two operations, there is still a race
> condition. I see no contradiction here.

On Posix systems, you can avoid the race condition. The trick is to
skip step #1. The rename will implicitly delete the old file, and
it will still be atomic. The whole process now consists of a single
stop, so the whole process is now atomic.

>> You don't need locks or any other fancy stuff. You just need to make
>> sure that you flush the data and metadata correctly to the disk and
>> force a re-write of the directory inode, too. It's a standard pattern on
>> POSIX platforms and well documented in e.g. the maildir RFC.

> It is not entirely true. We are talking about two processes. One is
> reading a file, another one is writting it. They can run at the same
> time, so flushing disk cache forcedly won't help.

On Posix systems, it will work, and be atomic, even if one process is
reading the old status file while another process is writing the new
one. The old file will be atomically removed from the directory by
the rename operation; it will continue to exists on the hard drive, so
that the reading process can continue reading it. The old file will
be deleted when the reader closes it. Or, if the system crashed before
the old file is closed, it will deleted when the system is restarted.

On Windows, things are very different.

Hope this helps,

-- HansM

[Ross Ridge]

Laszlo Nagy:
> This is not a contradiction. Although the rename operation is atomic,
> the whole "change status" process is not. It is because there are two
> operations: #1 delete old status file and #2. rename the new status
> file. And because there are two operations, there is still a race
> condition. I see no contradiction here.

Christian Heimes <li...@cheimes.de> wrote:
>Sorry, but you are wrong. It's just one operation that boils down to
>"point name to a different inode".

For some reason you're assuming POSIX semantics, an assumption that
Laszlo Nagy did not make.

Ross Ridge

--
l/ // Ross Ridge -- The Great HTMU
[oo][oo] rri...@csclub.uwaterloo.ca
-()-/()/ http://www.csclub.uwaterloo.ca/~rridge/
db //

[Laszlo Nagy]

> Sorry, but you are wrong. It's just one operation that boils down to

> "point name to a different inode". After the rename op the file name
> either points to a different inode or still to the old name in case of
> an error. The OS guarantees that all processes either see the first or
> second state (in other words: atomic).
>
> POSIX has no operation that actually deletes a file. It just has an
> unlink() syscall that removes an associated name from an inode. As soon
> as an inode has no names and is not references by a file descriptor, the
> file content and inode is removed by the operating system. rename() is
> more like a link() followed by an unlink() wrapped in a system wide
> global lock.

Then please help me understand this.

"Good" case:

process #1: unlink(old status file)
process #1: rename(new status file)
process#2: open(new status file)
process#2: read(new status file)

"Bad" case:

process #1: unlink(old status file)
process#2: open(???) -- there is no file on disk here, this system call
returns with an error!
process #1: rename(new status file)

If it would be possible to rename + unlink in one step, then it would be
okay. Can you please explain what am I missing?

[Laszlo Nagy]

>> This is not a contradiction. Although the rename operation is atomic,
>> the whole "change status" process is not. It is because there are two
>> operations: #1 delete old status file and #2. rename the new status
>> file. And because there are two operations, there is still a race
>> condition. I see no contradiction here.
> On Posix systems, you can avoid the race condition. The trick is to
> skip step #1. The rename will implicitly delete the old file, and
> it will still be atomic. The whole process now consists of a single
> stop, so the whole process is now atomic.

Well, I didn't know that this is going to work. At least it does not
work on Windows 7 (which should be POSIX compatible?)

>>> f = open("test.txt","wb+")
>>> f.close()
>>> f2 = open("test2.txt","wb+")
>>> f2.close()
>>> import os
>>> os.rename("test2.txt","test.txt")
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
WindowsError: [Error 183] File already exists
>>>

I have also tried this on FreeBSD and it worked.

Now, let's go back to the original question:

>>>This works well on Linux but Windows raises an error when status_file already exists.

It SEEMS that the op wanted a solution for Windows....

[Laszlo Nagy]

Yes, and actually it does not need to be an abnormal termination. It is
enough if the OS scheduler puts this process on hold for some time...

But using a lock file, the problem can be solved. However in that case,
reading a status file can be a blocking operation.

[Christian Heimes]

Am 12.07.2012 19:43, schrieb Laszlo Nagy:
> Well, I didn't know that this is going to work. At least it does not
> work on Windows 7 (which should be POSIX compatible?)

Nope, Windows's file system layer is not POSIX compatible. For example
you can't remove or replace a file while it is opened by a process.
Lot's of small things work slightly differently on Windows or not at all.

Christian

[Rick Johnson]

On Jul 12, 2:39 pm, Christian Heimes <li...@cheimes.de> wrote:
> Windows's file system layer is not POSIX compatible. For example
> you can't remove or replace a file while it is opened by a process.

Sounds like a reasonable fail-safe to me. Not much unlike a car
ignition that will not allow starting the engine if the transmission
is in any *other* gear besides "park" or "neutral", OR a governor (be
it mechanical or electrical) that will not allow the engine RPMs to
exceed a maximum safe limit, OR even, ABS systems which "pulse" the
brakes to prevent overzealous operators from loosing road-to-tire
traction when decelerating the vehicle.

You could say: "Hey, if someone is dumb enough to shoot themselves in
the foot then let them"... however, sometimes fail-safes not only save
the dummy from a life of limps, they also prevent catastrophic
"collateral damage" to rest of us.

[Steven D'Aprano]

On Thu, 12 Jul 2012 15:05:26 +0200, Christian Heimes wrote:

> You need to flush the data to disk as well as the metadata of the file
> and its directory in order to survive a system crash. The close()
> syscall already makes sure that all data is flushed into the IO layer of
> the operating system.

And some storage devices (e.g. hard drives, USB sticks) don't actually
write data permanently even when you sync the device. They just write to
a temporary cache, then report that they are done (liar liar pants on
fire). Only when the cache is full, or at some random time at the
device's choosing, do they actually write data to the physical media.

The result of this is that even when the device tells you that the data
is synched, it may not be.


--
Steven

[Chris Angelico]

On Fri, Jul 13, 2012 at 11:20 AM, Rick Johnson
<rantingri...@gmail.com> wrote:
> On Jul 12, 2:39 pm, Christian Heimes <li...@cheimes.de> wrote:
>> Windows's file system layer is not POSIX compatible. For example
>> you can't remove or replace a file while it is opened by a process.
>
> Sounds like a reasonable fail-safe to me.

POSIX says that files and file names are independent. I can open a
file based on its name, delete the file based on its name, and still
have the open file there. When it's closed, it'll be wiped from the
disk.

ChrisA

[Steven D'Aprano]

On Fri, 13 Jul 2012 12:12:01 +1000, Chris Angelico wrote:

> On Fri, Jul 13, 2012 at 11:20 AM, Rick Johnson
> <rantingri...@gmail.com> wrote:
>> On Jul 12, 2:39 pm, Christian Heimes <li...@cheimes.de> wrote:
>>> Windows's file system layer is not POSIX compatible. For example you
>>> can't remove or replace a file while it is opened by a process.
>>
>> Sounds like a reasonable fail-safe to me.

Rick has obviously never tried to open a file for reading when somebody
else has it opened, also for reading, and discovered that despite Windows
being allegedly a multi-user operating system, you can't actually have
multiple users read the same files at the same time.

(At least not unless the application takes steps to allow it.)

Or tried to back-up files while some application has got them opened. Or
open a file while an anti-virus scanner is oh-so-slooooowly scanning it.

Opening files for exclusive read *by default* is a pointless and silly
limitation. It's also unsafe: if a process opens a file for exclusive
read, and then dies, *no other process* can close that file.

At least on POSIX systems, not even root can override a mandatory
exclusive lock (it would be pretty pointless if it could), so a rogue or
buggy program could wreck havoc with mandatory exclusive file locks.
That's why Linux, by default, treats exclusive file locks as advisory
(cooperative), not mandatory.

In general, file locking is harder than it sounds, with many traps for
the unwary, and of course the semantics are dependent on both the
operating system and the file system.

https://en.wikipedia.org/wiki/File_locking

> POSIX says that files and file names are independent. I can open a file
> based on its name, delete the file based on its name, and still have the
> open file there. When it's closed, it'll be wiped from the disk.

One neat trick is to open a file, then delete it from disk while it is
still open. So long as your process is still running, you can write to
this ghost file, as normal, but no other process can (easily) see it. And
when your process ends, the file contents is automatically deleted.

This is remarkably similar to what Python does with namespaces and dicts:

# create a fake "file system"
ns = {'a': [], 'b': [], 'c': []}
# open a file
myfile = ns['a']
# write to it
myfile.append('some data')
# delete it from the "file system"
del ns['a']
# but I can still read and write to it
myfile.append('more data')
print(myfile[0])
# but anyone else will get an error if they try
another_file = ns['a']


--
Steven

[Gene Heskett]

On Thursday 12 July 2012 23:21:16 Steven D'Aprano did opine:

> On Fri, 13 Jul 2012 12:12:01 +1000, Chris Angelico wrote:
> > On Fri, Jul 13, 2012 at 11:20 AM, Rick Johnson
> >
> > <rantingri...@gmail.com> wrote:
> >> On Jul 12, 2:39 pm, Christian Heimes <li...@cheimes.de> wrote:
> >>> Windows's file system layer is not POSIX compatible. For example you
> >>> can't remove or replace a file while it is opened by a process.
> >>
> >> Sounds like a reasonable fail-safe to me.
>
> Rick has obviously never tried to open a file for reading when somebody
> else has it opened, also for reading, and discovered that despite
> Windows being allegedly a multi-user operating system, you can't
> actually have multiple users read the same files at the same time.
>

Chuckle. That was one of the 'features' that os9 on the trs-80 color
computer had back in the 80's, and it was clean and well done because of
the locking model the random block file manager had in OS9 for 6809 cpu's,
no relation to the Mac OS9 other than a similar name. That color computer
has a separate, text only video card I could plug in and display on an 80
column amber screen monitor.

When I wanted to impress the visiting frogs, I often did something I have
never been able to do on any other operating system since, start assembling
a long assembly language file on one of the screens on the color monitor,
hit the clear key to advance to the amber screen and start a listing on it
of the assemblers output listing file.

Because the file locking was applied only to the sector (256 bytes on that
machine) being written at the instant, the listing would fly by till it
caught up with the assemblers output, running into the lock and then
dutifully following along, one sector behind the assemblers output, until
the assembly was finished. That was in 1986 folks, and in the year of our
Lord 2012, 26 years later, I still cannot do that in linux. When I ask why
not, the replies seem to think I'm from outer space. Its apparently a
concept that is not even attempted to be understood by the linux code
carvers.

Something is drastically wrong with that picture IMO.

> (At least not unless the application takes steps to allow it.)
>
> Or tried to back-up files while some application has got them opened.

That in fact, ran me out of the amiga business in 1999, a 30Gb drive failed
on my full blown 040 + 64 megs of dram A2000. When the warranty drive
arrived is when I found that due to file locks on the startup files, all of
them involved with the booting of that machine, my high priced Diavolo Pro
backup tapes didn't contain a single one of those files. The linux box
with Red Hat 5.0 on it that I had built in late 1998 to see what linux was
all about found space under that desk yet that evening and I never looked
back.

Cheers, Gene
--
"There are four boxes to be used in defense of liberty:
soap, ballot, jury, and ammo. Please use in that order."
-Ed Howdershelt (Author)
My web page: <http://coyoteden.dyndns-free.com:85/gene> is up!
You just wait, I'll sin till I blow up!
-- Dylan Thomas

[Steven D'Aprano]

On Thu, 12 Jul 2012 23:49:02 -0400, Gene Heskett wrote:

> When I wanted to impress the visiting frogs, I often did something I
> have never been able to do on any other operating system since, start
> assembling a long assembly language file on one of the screens on the
> color monitor, hit the clear key to advance to the amber screen and
> start a listing on it of the assemblers output listing file.
>
> Because the file locking was applied only to the sector (256 bytes on
> that machine) being written at the instant, the listing would fly by
> till it caught up with the assemblers output, running into the lock and
> then dutifully following along, one sector behind the assemblers output,
> until the assembly was finished. That was in 1986 folks, and in the
> year of our Lord 2012, 26 years later, I still cannot do that in linux.

Um, what you are describing sounds functionally equivalent to what
tail -f does.

> When I ask why not, the replies seem to think I'm from outer space. Its
> apparently a concept that is not even attempted to be understood by the
> linux code carvers.

You could certainly create a pair of cooperative programs, one which
keeps a lock on only the last block of the file, and a tail-like reader
which honours that lock. But why bother? Just have the assembler append
to the file, and let people use any reader they like, such as tail.

Or have I misunderstood you?



--
Steven

[rantingri...@gmail.com]

On Thursday, July 12, 2012 10:13:47 PM UTC-5, Steven D&#39;Aprano wrote:
> Rick has obviously never tried to open a file for reading when somebody
> else has it opened, also for reading, and discovered that despite Windows

> being allegedly a multi-user operating system, you can&#39;t actually have

> multiple users read the same files at the same time.

You misread my response. My comment was direct result of Christian stating:

(paraphrase) "On some systems you are not permitted to delete a file whilst the file is open "

...which seems to be consistent to me. Why would *anybody* want to delete a file whilst the file is open? Bringing back the car analogy again: Would you consider jumping from a moving vehicle a consistent interaction with the interface of a vehicle? Of course not. The interface for a vehicle is simple and consistent:

1. You enter the vehicle at location A
2. The vehicle transports you to location B
3. You exit the vehicle

At no time during the trip would anyone expect you to leap from the vehicle. But when you delete open files, you are essentially leaping from the moving vehicle! This behavior goes against all expectations of consistency in an API -- and against all sanity when riding in a vehicle!

> Opening files for exclusive read *by default* is a pointless and silly

> limitation. It&#39;s also unsafe: if a process opens a file for exclusive

> read, and then dies, *no other process* can close that file.

Oh come on. Are you actually going to use "errors" or "unintended consequences", or even "Acts of God" to defend your argument? Okay. Okay. I suppose "IF" the car spontaneously combusted "THEN" the passengers would be wise to jump out, leaving the vehicle to the whims of inertia.

> One neat trick is to open a file, then delete it from disk while it is
> still open. So long as your process is still running, you can write to
> this ghost file, as normal, but no other process can (easily) see it. And
> when your process ends, the file contents is automatically deleted.

Well "neat tricks" aside, I am of the firm belief that deleting files should never be possible whilst they are open.

* Opening files requires that data exist on disk
* Reading and writing files requires an open file obj
* Closing files requires an open file object
* And deleting files requires that the file NOT be open

Would you also entertain the idea of reading or writing files that do not exist? (not including pseudo file objs like StringIO of course!).

Summary: Neat tricks and Easter eggs are real hoot, but consistency in APIs is the key.

[Chris Angelico]

On Fri, Jul 13, 2012 at 2:26 PM, <rantingri...@gmail.com> wrote:
> On Thursday, July 12, 2012 10:13:47 PM UTC-5, Steven D&#39;Aprano wrote:
>> Rick has obviously never tried to open a file for reading when somebody
>> else has it opened, also for reading, and discovered that despite Windows
>> being allegedly a multi-user operating system, you can&#39;t actually have
>> multiple users read the same files at the same time.
>
> You misread my response. My comment was direct result of Christian stating:
>
> (paraphrase) "On some systems you are not permitted to delete a file whilst the file is open "
>
> ...which seems to be consistent to me. Why would *anybody* want to delete a file whilst the file is open?

POSIX doesn't let you delete files. It lets you dispose of filenames.
Python does the same with its 'del'. The object (file) exists until
the system decides otherwise.

Here's a simpler example: Hardlinks. Suppose you have two names
pointing to the same file; are you allowed to unlink one of them while
you have the "other" open?

ChrisA

[Steven D'Aprano]

On Thu, 12 Jul 2012 21:26:20 -0700, rantingrickjohnson wrote:

> On Thursday, July 12, 2012 10:13:47 PM UTC-5, Steven D&#39;Aprano wrote:
>> Rick has obviously never tried to open a file for reading when somebody
>> else has it opened, also for reading, and discovered that despite
>> Windows being allegedly a multi-user operating system, you can&#39;t
>> actually have multiple users read the same files at the same time.
>
> You misread my response. My comment was direct result of Christian
> stating:
>
> (paraphrase) "On some systems you are not permitted to delete a file
> whilst the file is open "
>
> ...which seems to be consistent to me. Why would *anybody* want to
> delete a file whilst the file is open?

Because it is useful and a sensible thing to do.

Why should one misbehaved application, keeping a file open, be allowed to
hold every other application, and the file system, hostage?

This is one of the many poor decisions which makes Windows so vulnerable
to viruses and malware. If malware can arrange to keep itself open, you
can't delete it. Thanks guys!

> Bringing back the car analogy
> again: Would you consider jumping from a moving vehicle a consistent
> interaction with the interface of a vehicle? Of course not. The
> interface for a vehicle is simple and consistent:
>
> 1. You enter the vehicle at location A
> 2. The vehicle transports you to location B
> 3. You exit the vehicle

Amusingly, you neglected to specify "the vehicle stops" -- and rightly
so, because of course having to stop the vehicle is not a *necessary*
condition for exiting it, as tens of thousands of stunt men and women can
attest.

Not to mention people parachuting out of an airplane, pirates or
commandos boarding a moving ship, pedestrians transferring from a slow
moving walkway to a faster moving walkway, farmers jumping off a trailer
while it is still being towed behind a tractor (and jumping back on
again), and Bruce Willis in "Red" in very possibly the best slow-motion
action sequence in the history of Hollywood.

http://www.youtube.com/watch?v=xonMpj2YyDU

> At no time during the trip would anyone expect you to leap from the
> vehicle.

Expected or not, you can do so.

> But when you delete open files, you are essentially leaping
> from the moving vehicle! This behavior goes against all expectations of
> consistency in an API -- and against all sanity when riding in a
> vehicle!

Fortunately, files on a file system are not cars, and deleting open files
is a perfectly reasonable thing to do, no more frightening than in Python
deleting a reference to an object using the del statement. Imagine how
stupid it would be if this happened:


py> x = 42
py> y = x
py> del y

Traceback (most recent call last):
File "<stdin>", line 1, in <module>

DeleteError: cannot delete reference to object '42' until no other
references to it exist


Fortunately, Python doesn't do that -- it tracks when the object is no
longer being accessed, and only then physically reclaims the memory used.
And so it is on POSIX file systems: the file system keeps track of when
the file on disk is no longer being accessed, and only then physically
reclaims the blocks being used. Until then, deleting the file merely
unlinks the file name from the blocks on disk, in the same way that
"del y" merely unlinks the name y from the object 42.

>> Opening files for exclusive read *by default* is a pointless and silly
>> limitation. It&#39;s also unsafe: if a process opens a file for
>> exclusive read, and then dies, *no other process* can close that file.
>
> Oh come on. Are you actually going to use "errors" or "unintended
> consequences", or even "Acts of God" to defend your argument?

Features have to be judged by their actual consequences, not some
unrealistic sense of theoretical purity. The actual consequences of
mandatory exclusive file locking is, *it sucks*.

Windows users are used to having to reboot their server every few days
because something is broken, so they might not mind rebooting it because
some file is locked in a mandatory open state and not even the operating
system can unlock it. But for those with proper operating systems who
expect months of uninterrupted service, mandatory locking is a problem to
be avoided, not a feature.

> Okay.
> Okay. I suppose "IF" the car spontaneously combusted "THEN" the
> passengers would be wise to jump out, leaving the vehicle to the whims
> of inertia.

In this analogy, is the car the file name, the inode, or the directory?
Are the passengers the file name(s), or the file contents, or the inode?
Is the driver meant to be the file system? If I have a hard link to the
file, does that mean the passengers are in two cars at once, or two lots
of passengers in the same car?

>> One neat trick is to open a file, then delete it from disk while it is
>> still open. So long as your process is still running, you can write to
>> this ghost file, as normal, but no other process can (easily) see it.
>> And when your process ends, the file contents is automatically deleted.
>
> Well "neat tricks" aside, I am of the firm belief that deleting files
> should never be possible whilst they are open.

[condescension = ON]

Good for you Rick. Having strongly held opinions on things you have only
a limited understanding about is your right as an American.

[condescension = OFF]

> * Opening files requires that data exist on disk
> * Reading and writing files requires an open file obj

You have missed a step in jumping from files on disk to open file objects.

Open file objects do not necessarily correspond to files on disk. For
example, in standard Pascal, file objects are purely in-memory constructs
emulating files on a tape drive, with no relationship to on-disk files.

(Any half-decent Pascal compiler or interpreter will *also* give you ways
to access real files on disk, but that isn't covered by the standard.)

Even when the file object does come from an actual disk file, we can
conclude that before you can open a file for reading, it must exist; but
having opened it, there is no *necessary* requirement that it *remains*
on disk. If you try to read from an open file object whose underlying
file has been deleted, there are three perfectly reasonable behaviours:

- you get an error;

- it is the same result as if the file has been truncated to zero bytes;

- deleting the file only deletes the *name*, not contents, until the
last open file handle is shut, and then the contents are deleted.

> * Closing files requires an open file object

Naturally; but open file objects don't require that the on-disk file
still exists.

> * And deleting files requires that the file NOT be open

Not at all.

> Would you also entertain the idea of reading or writing files that do
> not exist? (not including pseudo file objs like StringIO of course!).

Define "file" and "exist". Because you are conflating at least three
different things:

a file name
an inode (blocks on a disk)
a file object

Of course it is useful to break the abstraction that file objects must be
files on disk. StringIO is one such example. Standard Pascal file objects
is another. Likewise, it is useful to be able to read from an inode that
is no longer connected to a file name.

So, absolutely, yes, it is useful to be able to read and write from files
that don't exist, under some circumstances.

> Summary: Neat tricks and Easter eggs are real hoot, but consistency
> in APIs is the key.

A foolish consistency is the hobgoblin of little minds.


--
Steven

[Prasad, Ramit]

> Well "neat tricks" aside, I am of the firm belief that deleting files should
> never be possible whilst they are open.

This is one of the few instances I think Windows does something better
than OS X. Windows will check before you attempt to delete (i.e. move
to Recycling Bin) while OS X will move a file to Trash quite happily
only tell me it cannot remove the file when I try to empty the Trash.

Ramit


Ramit Prasad | JPMorgan Chase Investment Bank | Currencies Technology
712 Main Street | Houston, TX 77002
work phone: 713 - 216 - 5423

--


This email is confidential and subject to important disclaimers and
conditions including on offers for the purchase or sale of
securities, accuracy and completeness of information, viruses,
confidentiality, legal privilege, and legal entity disclaimers,
available at http://www.jpmorgan.com/pages/disclosures/email.

[Chris Gonnerman]

On 07/13/2012 11:00 AM, Prasad, Ramit wrote:
>> Well "neat tricks" aside, I am of the firm belief that deleting files should
>> never be possible whilst they are open.

> This is one of the few instances I think Windows does something better
> than OS X. Windows will check before you attempt to delete (i.e. move
> to Recycling Bin) while OS X will move a file to Trash quite happily
> only tell me it cannot remove the file when I try to empty the Trash.

While I was trained in the Unix way, and believe it is entirely
appropriate to delete an open file. Even if I my program is the opener.
It's just too handy to have temp files that disappear on their own.

As opposed to periodically going to %TEMP% and deleting them manually. Gah.

-- Chris.

[Prasad, Ramit]

> >> Well "neat tricks" aside, I am of the firm belief that deleting files
> should
> >> never be possible whilst they are open.

> > This is one of the few instances I think Windows does something better
> > than OS X. Windows will check before you attempt to delete (i.e. move
> > to Recycling Bin) while OS X will move a file to Trash quite happily
> > only tell me it cannot remove the file when I try to empty the Trash.
> While I was trained in the Unix way, and believe it is entirely
> appropriate to delete an open file. Even if I my program is the opener.
> It's just too handy to have temp files that disappear on their own.
>
> As opposed to periodically going to %TEMP% and deleting them manually. Gah.

In my experience things that are "too handy" are usually breaking
what I consider "right". That being said, I am not entirely sure
what I think is "right" in this circumstance. I suppose it depends
on if I am the person deleting or the person who is looking at
a file that is being deleted. Or the user who just wants the stupid
computer to just Work.

I lean slightly towards the POSIX handling with the addition that
any additional write should throw an error. You are now saving to
a file that will not exist the moment you close it and that is probably
not expected.





Ramit

[Chris Angelico]

On Sat, Jul 14, 2012 at 3:59 AM, Prasad, Ramit
<ramit....@jpmorgan.com> wrote:
> I lean slightly towards the POSIX handling with the addition that
> any additional write should throw an error. You are now saving to
> a file that will not exist the moment you close it and that is probably
> not expected.

There are several different possible "right behaviors" here, but they
depend more on the application than anything else. With a log file,
for instance, the act of deleting it is more a matter of truncating it
(dispose of the old history), so the right thing to do is to start a
fresh file. Solution: Close the file and re-open it periodically. But
I don't know of an efficient way to do that with Windows semantics.
Renaming/moving an open file in order to perform log rotation isn't
all that easy.

ChrisA

[Hans Mulder]

On 13/07/12 19:59:59, Prasad, Ramit wrote:

> I lean slightly towards the POSIX handling with the addition that
> any additional write should throw an error. You are now saving to
> a file that will not exist the moment you close it and that is
> probably not expected.

I'd say: it depends.

If the amount of data your script needs to process does not fit
in RAM, then you may want to write some of it to a temporary file.
On a Posix system, it's entirely normal to unlink() a temp file
first thing after you've created it. The expectation is that the
file will continue to exists, and be writeable, until you close it.

In fact, there's a function in the standard library named
tempfile.TemporaryFile that does exactly that: create a file
and unlink it immediately. This function would be useless
if you couldn't write to your temporary file.

[MRAB]

On 13/07/2012 19:28, Hans Mulder wrote:
> On 13/07/12 19:59:59, Prasad, Ramit wrote:
>
>> I lean slightly towards the POSIX handling with the addition that
>> any additional write should throw an error. You are now saving to
>> a file that will not exist the moment you close it and that is
>> probably not expected.
>

Strictly speaking, the file does exist, it's just that there are no
names referring to it. When any handles to it are also closed, the file
_can_ truly be deleted.

As has been said before, in the *nix world, "unlink" _doesn't_ delete
a file, it deletes a name.

> I'd say: it depends.
>
> If the amount of data your script needs to process does not fit
> in RAM, then you may want to write some of it to a temporary file.
> On a Posix system, it's entirely normal to unlink() a temp file
> first thing after you've created it. The expectation is that the
> file will continue to exists, and be writeable, until you close it.
>
> In fact, there's a function in the standard library named
> tempfile.TemporaryFile that does exactly that: create a file
> and unlink it immediately. This function would be useless
> if you couldn't write to your temporary file.
>

It's possible to create a temporary file even in Windows.

[Chris Gonnerman]

On 07/13/2012 12:59 PM, Prasad, Ramit wrote:
> I lean slightly towards the POSIX handling with the addition that any
> additional write should throw an error. You are now saving to a file
> that will not exist the moment you close it and that is probably not

> expected. Ramit
But if I created, then deleted it while holding an open file descriptor,
it is entirely likely that I intend to write to it. I'll admit, these
days there are those in the Unix/Linux community that consider using an
anonymous file a bad idea; I'm just not one of them.

-- Chris.

[Christian Heimes]

Am 13.07.2012 21:57, schrieb MRAB:
> It's possible to create a temporary file even in Windows.

Windows has a open() flag named O_TEMPORARY for temporary files. With
O_TEMPORARY the file is removed from disk as soon as the file handle is
closed. On POSIX OS it's common practice to unlink temporary files
immediately after the open() call.

[Steven D'Aprano]

A badly-behaved application can write oodles and oodles of data to an
unlinked file, which has the result of temporarily using up disk space
that doesn't show up when you do an ls. For an inexperienced system
administrator, this may appear mysterious.

The solution is to us lsof to identify the unlinked file, which gives you
the process id of the application, which you can then kill. As soon as
you do that, the space is freed up again.

Like all powerful tools, unlinked files can be abused. Underpowered tools
can't be abused, but nor can they be used.


--
Steven

[Christian Heimes]

Am 13.07.2012 03:52, schrieb Steven D'Aprano:
> And some storage devices (e.g. hard drives, USB sticks) don't actually
> write data permanently even when you sync the device. They just write to
> a temporary cache, then report that they are done (liar liar pants on
> fire). Only when the cache is full, or at some random time at the
> device's choosing, do they actually write data to the physical media.
>
> The result of this is that even when the device tells you that the data
> is synched, it may not be.

Yes, that's another issue. Either you have to buy expensive enterprise
hardware with UPS batteries or you need to compensate for failures on
software level (e.g. Hadoop cluster).

We have big storage devices with double redundant controllers, on board
buffer batteries, triple redundant power supplies, special RAID disks,
multipath IO fiber channel links and external backup solution to keep
our data reasonable safe.

Christian