Slashdot Mirror
Improving Unix Mail Storage?
At first, there was mbox, then there was Maildir, and Bill begat Outlook and .mbx. CaraCalla wonders if there is a better way to store mail than the way we currently store it today. I admit, with the changes that email has undergone over the past 5 years (changes in what is being sent, not necessarily in how it is sent), it may be time to reinvent the mail format. Read on for CaraCalla's analysis of the current mail options, and his thoughts on where we may go in the future. If you were to design your own MUA, how would you design its mail storage?
CaraCalla asks: "Does anybody know a good, free solution for storing mail on unix hosts? The reason that I ask this question is my discontent with available techniques:
- mbox: There are problems with locking, corruption, access-times, and bloat.
- Maildir: Do you really want to clutter your system with millions of small files? That's waste of inodes, space (unless perhaps you use Linux/ReiserFS or SGi) and just try to open a Maildir with 1000+ mails and see how long it takes your favorite Mailprogram to only display the subjects.
- Cyrus: Basically the same as Maildir with database features.
- UW-Imap mbx: That's classical mbox with extensions allowing multiple access.
- Evolution: Basically mbox with database features.
- Windows clients: Typically some proprietary db-format. Pathetic.
But the thing that bugs me most is disk space. Typical inboxes are made of 5% to 10% of Text including Headers and HTML. The rest are BASE64- (or UU-) encoded pictures, word documents, zip archives and so on. The problem here is the encoding which wastes considerable amounts of space (at least one third).
Some ideas about the ideal mail-storage:
- One file per Mailbox-folder, allowing multiple folders per user. Should those files reside in one central location or in users Homedirs?
- Compression: Should messages be broken into pieces and the MIME-attachments stored separately (thus searching of the text parts would still be possible without decompressing the whole file)?
- File format: gdbm, Sleepycat db? Something new?
- Should the security model allow users to directly access their files, grep them, copy them around?
- Shared folders, virtual domains?
- Unicode support in folder names? Imap message-IDs, flags, useragent specific state-information?
- How would MTAs deliver mail? How would clients access? File-locking (NFS)?
- What about backwards-compatibility? Writing libmailstore (anyone)? adopting UW c-client?
Does my ideal mailstorage exist somewhere? Is somebody working on a project addressing this? Does anybody have some other hints? And please no mbox/Maildir flamewar!"
But put multiple indexes (by sender, subject, date, whatever key-classes you want to assign messages) and the possibility to restrict the range displayed. With careful programming, you can manage many users who won't be able to read each other's mail, except as required.
This way, you can arrange your mail as you please.
No more message duplication. Send a memo to 250 people? Just send it once, but tag it as readable by the 250 sendees.
Of course, this calls for an SQL database... :) :) :)
Plain and simple. Switch from mail to Usenet. Maildir-like structure, but with a .overview (XOVER) file to help out with indexing.
Storage is another problem, though... but Usenet messages can be sidetracked a bit with the encoding.
--
# Canmephians for a better Linux Kernel
$Stalag99{"URL"}="http://stalag99.net";
I lost the plot half way through this, but here's some food for thought anyway. Now I should get back to work ...
;), but it works reasonably well, and we've had a chance to try and deal with users with lots of email.
...
...
...
... and you dont want 1-tier applications), so it doesn't matter what format you use under the belt - you can choose the format which best suits what you're trying to do.
Z
I think that this is looking for the solution to a problem that doesn't really exist in the first place. Although I guess it depends somewhat on what you define as 'Unix mail'.
I'm a developer on Evolution, and primarily on Camel, evolution's email library. I'm not sure i'd rave about it (although I think Camel is a mostly beautiful piece of code
What IS 'Unix mail'?
I would define Unix mail as mail (rfc822 format) downloaded and stored locally on a per-user basis. IMAP, Exchange, and other remote protocols are very different beasts.
Why are DBMS's not suitable for 'Unix mail'?
Once you have a remote server you have to do things differently than if you have local access. Using a DBMS, and having a trained administrator to manage it are practical considerations, as are the benefits you might get from this configuration. These solutions dont really make sense for standalone users. They shouldn't need to install and manage databases, complex backup prodedures, and so forth, just to read their email.
i.e. rdbms's are:
hard to setup
hard to maintain
another major point of failure
If however, I was to design a multi-user groupware server, then a DBMS would come into serious consideration - at the backend at least. It allows you do to things like easily consolidate authentication outside of the operating system (the idea of having a 'shell account' to access mail is somewhat outdated), it allows you to save space by storing common data, like attachments and email content in a single place, and redirecting it to multiple recipients (which is a common practice within organisations). It may be practical to use a mixture, a RDBMS to store textual parts or indices to data stored in a more conventional filesystem.
But even with a RDBMS backend, I would personally probably still stick to IMAP to serve it to actual clients. The IMAP protocol is a bit heavy, but not really that bad, and it serves email, I dont think there's really any need to reinvent the wheel here.
So
If you define unix mail as I have, and separate it from a *mail server*, then you rule out full blown RDBMS's, and are left with:
single file database
multiple file database
I'm not even going to mention XML because I think it is the single most stupid idea anyone's come up with. It is completely unsuitable for this purpose.
And well, there's really no reason not to use MIME to store the messages. MIME already does everything you can possibly do with email (since, uh, it is how the email *will* be sent), any client will already have to deal with it, and mime decoding is for the most part really quite simple and fast anyway. Translating the mime format into some other storage format really doesn't make sense.
single file databases
mbox
Mbox is a single file database. Its just that everyone that uses it generally writes their own access code. This is where problems with 'locking' come about, either because the underlying filesystem doesn't support it properly (e.g. some nfs implementations), or everyones clients don't use the same locking mechanism. This really just an implementation issue anyway. There would be nothing to stop someone writing a common 'mbox.db' library that stored everything in completely compatible mbox files, which took all the work out of it, and then you'd have an mbox DBMS
mbox scales ok, without any caching of header information it handles in the order of 2K messages in an interactive timescale, and quite a lot more if you dont mind some short delays (i.e. in the order of the time it takes mozilla to start up).
Appending and reading is quick, and reliable - assuming the filesystem works, which is a pretty safe assumption to make. This is assuming the mailbox is first summarised at first opening, otherwise looking up messages can be slow, because you have to scan the whole file first.
The only operation that is slow is expunging messages, and at worst case isn't really any slower than copying a whole file across to another file.
The only other issue is agreement on the 'standard' for what constitutes an mbox file. For example. Solaris uses and honours the 'Content-Length' header, and thus it does not translate any lines beggining with "From " into the conventional ">From ". Some mail clients translate "(>*)From " into ">\1From " (using sed syntax) and visa versa, others do not. There is no standard, just some conventions, some of which aren't easy to determine either.
Because you need to keep the whole index in memory at once, this can become expensive, but you could use a secondary database as an index into the real file. But eventually you hit a point where the cost of expunging does get too expensive. You could just archive the mail regularly, or use a format like maildir instead.
gdbm/db/etc
db files wrap the single file in a common api that handles all of the locking issues and access issues for you. Some have different features, e.g. querying capability, logging and transactions, etc.
We've never tried to use db for this purpose, more just because we didn't think it was worth it. All you really get with a minimal implementation is the ability to store and retrieve a blob of data using a single key. Writing is fairly slow because the database has to manage more details for you (locking, allocating blocks, unlocking, etc). You could use multiple db files as indices to perform multiple-key searches, but they are quite slow at creating them (we tried using db for the content indices and it was way too slow).
i.e. even if you store the data in a db file, which gives you a slight benefit of inbuilt referential integrity, you still need to provide additional indices to actually be able to use it in any useful way. Evolution suffers this problem with the addressbook which stores vCards in db records.
Most db libraries (all?) also dont provide any mechanism to stream data. You either get the whole lot into memory, or you get none of it. So for large messages you're limited by memory (well, evolution is anyway, but it doesn't have to be). Yes, memory is cheap, but it is still a consideration, and it would certainly rule out a simple database in a multi-user environment.
db files are also slower than native files, especially for large objects. You're mapping an arbitrarily sized chunk of data to some 'database blocks', which are then stored in an arbitrarily sized 'database file' which the operating system is then mapping to its 'filesystem blocks'.
multifile solutions
Well I guess this comes down to mh and maildir. mh isn't really suitable for anything, because of its just plain bad design and lack of defined semantics. There's no way to guarantee anything about its operation.
maildir - i like. It moves the scourge of trying to implement a reliably, scalable, multiple access database almost entirely into the operating system layer. Operating systems already do this very well - they manage hundreds of thousands of files randomly written across your disks, without skipping a beat.
No operation requires more than a single message size of data, and the operating system already indexes the message, via its filename. Sure, ext2 doesn't do such a swell job with long directories, but that can be addressed (and the same problem can be addressed on just about any platform). For 'free' you get concurrent multiple-reader, multiple-writer database access, without any of the considerable problems you have to solve to implement it otherwise.
The maildir 'protocol' is simple, reliable, and it works.
Again, it can easily be augmented by a client with additional indices, but for things like delivery agents who dont care about existing email, they dont need to suffer that overhead at all.
Some other comments specific to the question:
Compression. Personally I dont see the point. But a maildir-like structure would fit well with compression. Flat files would be the worst (e.g. mbox), and block-file formats (like db files) would also work well with compression. The good thing about email is it is 'write once', you don't edit or change the messages in the mailbox.
External attachments. I guess its possible, but again, it isn't really worth it in most cases. Parsing MIME is *fast*. It is much faster than parsing xml, and besides, people rarely look at an email more than once or twice. There isn't much use going off and storing the attachment in a high-performance reading format if it isn't going to be accessed often, and it just places a greater burden on your server.
base64, etc. Well, its entirely possible simply to store the messages as 'binary' format. Assuming the boundary markers are checked properly, Camel can work with binary encoded mail messages, and probably at least some other mail clients can too. There are some problems with some of the extremely broken openpgp/pgp/mime specs which suddenly say that mail transports aren't allowed to alter the *transport* encodings of some parts, but well, these specs are just braindead, and can be worked around.
Security model. Well, talking about Unix mail, not server mail, the filesystem is adequate.
Shared folders - is not an issue for unix mail.
Unicode. Well you can write unicode filenames to most unix filesystems, evne if 'ls' doesn't show it right.
MTA. Nothing could be simpler or safer than maildir as a delivery format. The mta doesn't have to care about any client-side indices, the mua will simply update them when it incorporates the new messages, etc.
Writing libmailstore? Mate, its called Camel, and its already written. Camel already does mbox, maildir, mh, it can read spool files directly (it doesn't create a summary file or build any indexes), it can talk imap, pop, and partial support for nntp. If someone gave me a decent RDBMS table schema and a carton of pale, I could probably write a MySQL backend in a couple of days, well, assuming the MySQL api is mt-safe.
Finally, some comments on evolution.
Evolution isn't reinventing any wheel. We use standard mbox format (if such a thing really exists anyway). We use standard maildir format, etc. Yes we may optionally create body indices, and we do usually create on-disk binary/compressed 'summaries' of the data, but these are really just on-disk caches of in-memory data structures, rather than anything to do with the mail storage format.
We put mail in another location, but everyone else has done that too, elm:Mail, pine:mail (or is it the other way around?), netscape:ns_mail, etc. At least we now offer the option to read most of this 'in place'.
The main problems evolution has with scalability is:
indexing.
Indexing is quite costly. The original index code was written somewhat like a database, it handled all internal data structures, used blocks of data, etc. It was slow, it scaled poorly. Definetly some of the algorithm choices and the implementation wasn't that hot, but it shows that such a solution isn't as simple as at first thought. Using libdb was impossibly slow (like several orders of magnitude slower).
The new stuff is a lot better, but can still use a lot of resources while indexing, and copies the whole file (well 2 files) across when performing expunges, but they are only performed occasionally, and the indices are smaller than the original indices, so in practice it scales much much better.
the summaries
The summaries are indices of a sort anyway. They are an in-memory tree of a subset of the information on each message. Enough information to display a list of messages, and perform vfoldering operations. Even though we do some tricks, like sharing common strings, the summary can get very large.
But, its a tradeoff I thought was worth it, rather than using on-disk summaries. The api's are much easier to use, and the problem gets pushed to the user - if they want to have folders with 100K messages, they should expect it to use a bit of memory. The on-disk size of the summaries is very small too, although I guess it could be made even smaller if we consolidated common strings.
per-message memory use
Currently, a lot of data gets copied around in memory. Every time you read a message, at least 1 whole copy of the (decoded) message is in memory at a given time (yes, including attachments). For IMAP this can get even worse (2-3 copies of a given attachment at a given time), because it doesn't stream enough. Most of this could use a disk-backing without changing any api's though, and well, i'm rewriting IMAP.
Wrapping up
And yeah, we're talking 100K messages here, not 1400. My 500Mhz celeron laptop has about 35K messages stored over about 10 mbox files, and it starts up in under 10 seconds, and that includes all of the bonobo/activation overhead (which is very significant). Yeah it uses a bit of memory, but memory is cheap on a personal workstation.
In short. The current mailbox formats we have suffice for "Unix mail". Add some archiving abilities to your mail client (even RDBMS backed mail clients need archiving), and you'll never have to delete a message again, and still get work done and still use mbox.
If you want to talk about writing a server - well who cares, you can do whatever you want, because everyone has to go through your interface anyway (you DO NOT want clients accessing data under you, thats what DBMS's are all about in the first place
It seems some people think using 1-tier applications (client code talking directly to a database) are the way to go for multi-user environments. They're not, they dont scale and are impossible to maintain. Nobody writes any real software like that anymore, unless you're writing dodgey vb toy apps.
_
\\/ are accustomed' - First Lensman
here's a little transcript:
% cd /mail/fs/mbox
/mail/fs/mbox/318/2/body is a jpeg file, viewable directly by any usual jpeg viewer).
% lc
Directories:
1 113 128 142 157 171 186 20 214 229 243 258 272 287 300 315 33 344 359 373 388 401 416 430 445 46 474 56 70 85
[...]
% cd 318
% lc
Files:
bcc date filename info messageid rawbody sender type body digest from inreplyto mimeheader rawheader subject unixheader cc disposition header lines raw replyto to
Directories:
1 2 3
% head raw
Return-Path:
Received: from punt-1.mail.demon.net by mailstore for rog@vitanuova.com
id 1021665470:10:17045:138; Fri, 17 May 2002 19:57:50 GMT
Received: from psuvax1.cse.psu.edu ([130.203.4.6]) by punt-1.mail.demon.net
id aa1016828; 17 May 2002 19:57 GMT
Received: from psuvax1.cse.psu.edu (psuvax1.cse.psu.edu [130.203.6.6])
by mail.cse.psu.edu (CSE Mail Server) with ESMTP
id 27DA4199BE; Fri, 17 May 2002 15:57:13 -0400 (EDT)
Delivered-To: 9fans@cse.psu.edu
Received: from acl.lanl.gov (plan9.acl.lanl.gov [128.165.147.177])
% head body
This is a multi-part message in MIME format.
--upas-mbyuptynpdsmbjuyeermihdgur
Content-Disposition: inline
Content-Type: text/plain; charset="US-ASCII"
Content-Transfer-Encoding: 7bit
Hi,
If you seek excitement and thrills you need to look no further than
Plan9 -- it gives you everything and then some, but in a good way (or
% cd 2
% lc
Files:
bcc date filename info messageid rawbody sender type
body digest from inreplyto mimeheader rawheader subject unixheader
cc disposition header lines raw replyto to
% cat mimeheader
Content-Type: image/jpeg
Content-Disposition: attachment; filename=iostats.jpg
Content-Transfer-Encoding: base64
% page body
reading through graphics...
%
"raw" contains the raw data that makes up the message. "body" contains the data after the encoding formats have been applied (hence in that case
the beauty of this scheme is that it hides the underlying storage scheme from the mail clients. if i wish to change things so that the underlying storage format is many files [currently it uses a traditional mbox format], none of the mail client programs have to change.
plus i can use grep, diff, shell scripts, etc directly on the messages in my mailbox. procmail eat your heart out.