[Zope-dev] dtml-sendmail and transactions: Request For Comments

[Toby Dickenson]

On Fri, 10 Aug 2001 22:43:59 +0200 (CEST), Dieter Maurer
<dieter@handshake.de> wrote:

>Good, but isn't it difficult to make mail sending transactional:
>
>    The sending may raise exceptions.
>
>    Zope is very angry to get exceptions during the second
>    phase of the two-phase commit.

http://dev.zope.org/Wikis/DevSite/Proposals/TransactionalEmail
outlines how I plan to minimize this risk. 99% of the sending work is
done during the rendering of the dtml-sendmail tag; it does everything
up to sending the final "." which indicates the end of the message
body.

This allows the most common errors to be detected in the same place as
today. These errors include:
1. failure to connect
2. server configuration prevents email relaying
3. destination address wrong (for some types of wrong)

>    This means, sending must go into the first phase.
>    But in this case, they are sent even if a different
>    transaction manager rejects the "commit" request
>    (or raises an exception).

I plan to send the final "." in the second phase, only after all other
transactional components have agreed that the transaction should
complete.

At this point there is very little that can go wrong: the mail server
running out of disk space is top of the list. I plan to convert any
exceptions into log messages (because, as you say, Zope would be angry
for me to let the exception escape).

I dont see this as a major problem in most cases. (unless we are too
far out of disk space to store the log message too ;-(


On Sat, 11 Aug 2001 12:29:32 -0500, "Phillip J. Eby"
<pje@telecommunity.com> wrote:

>The way Ty and I handle this in our applications is to not use the =
sendmail=20
>tag, but instead write a file to a queue directory, where it is then =
sent=20
>by another process which monitors the queue.  This is easier to make=20
>transactional, since the commit operation only has to move the file to=20
>commit it.
>
>Unfortunately, this is not a general solution for Zope, since it then=20
>requires the existence of that other process.  But in theory one could=20
>create a Product that would implement a daemonic thread for sending =
queued=20
>mail in this fashion.

It would be nice if this other process could deliver mail directly to
its recipients, without having to go via another mailer process.=20

An alternative way of looking at the problem; it would be nice if
sendmail could grow a two-phase commit extension.


Toby Dickenson
tdickenson@geminidataloggers.com