Re: [PATCH RFC 1/6] send-email: Add --delay for separating emails

On Tue, Apr 07, 2009 at 04:25:17PM -0500, Michael Witten wrote:

quoted

When sending a patch series, the emails often arrive at the final
destination out of order; though these emails should be chained
via the In-Reply-To headers, some mail-viewing systems display
by order of arrival instead.

The --delay option provides a means for specifying that there
should be a certain number of seconds of delay between sending
emails, so that the arrival order can be controlled better.

On Tue, Apr 7, 2009 at 16:51, Jeff King wrote:

I'm a little dubious how well this works in practice. Have you done
any experiments?

I have indeed, and I got better results. Whether that was a fluke,
I can't say for sure.

[Delivery delays] can be much larger than a few seconds, and this
won't help at all there.

and

On Tue, Apr 7, 2009 at 18:17, Junio C Hamano wrote:

I do not think giving N second interval at the sending end would
help much in the real world. Between your submitting MUA (that's
"git-send-email") and the client MUA, there are many hops involved...
any single hop can batch the messages that arrive within a small time
window before passing them to the next hop, and it can reorder the
messages when it does so.

In short, the only thing your --delay can control is the arrival
interval at your outgoing MSA. The arrival interval and order of
messages are outside your control for later hops.

For the most part, yes, I am operating under the assumption that email
is sent as soon as it is received by any intervening hop (no batch
accumulation). However, I'm also postulating that there always exists
an N that serves as an upperbound on the transit time, regardless of
batching.

That's where these comment comes into play:

On Tue, Apr 7, 2009 at 16:51, Jeff King wrote:

The reason I am dubious is that you are presumably delaying only a few
seconds (since anything more would be quite annoying to the user).

On Wed, Apr 8, 2009 at 01:03:

A multi-second delay is downright annoying. As a sender, I don't think
I would enable this option.

However, this sentiment doesn't make sense to me.

Firstly, I presume that someone is electing to use this option, so it is
almost by definition not annoying for that person.

Secondly, it seems reasonable to drop into another VC, screen window, or
terminal instance, and then set send-email a-running. For instance, with
a 14-patch series, one could set `--delay 60' and then let send-email
run happily for the next 14 minutes with nary a thought.

However, I think you're coming at it from a different angle:

On Wed, Apr 8, 2009 at 01:03, Jeff King wrote:

But apparently many readers sort by date received. See this subthread:

 http://article.gmane.org/gmane.comp.version-control.git/110097

I am generally of the opinion that if it is a big problem for people,
they should get a better mail client. But I am also open to suggestions
for helping receivers on crappy mail clients as long as those
suggestions do not put a burden on the sender.

and

on Tue, Apr 7, 2009 at 17:08, Nicolas Sebrecht wrote:

If the receiver wants the patch series be in a good ordered _for
sure_, he has to switch to a client mail supporting the In-Reply-To
chains.

Frankly, I don't care how other people's patch series appear to me. I care
about how mine appear to others. Is this irrational? Probably, but I'm kind
of OC; I want my patch series to look like it's in order for everyone.

This has nothing to do with what the receiver wants. This has everything
to do with what the sender wants. I want my patch series to be in order
even for wrongheaded receivers.

So, I never had any intention of forcing a delay on the sending end. It is
strictly for the sending end to use if desired.

On Tue, Apr 7, 2009 at 18:17, Junio C Hamano wrote:

I think send-email already has hacks to timestamp the messages at
least one-second apart by shifting the Date: field, so that the
recipient MUA can sort by the departure timestamp if it wants to (and
if it can), instead of the arrival timestamp. Is it not working well
for you?

I have worked with 2 major mail clients that both display by date received:

  * Mac OS X's Mail
  * Gmail

I assume that a not-insignificant number of others also use these clients.
I don't mind letting send-email run in the background for a few minutes if
it means my patch series appear in order.

As you say, Jeff:

On Wed, Apr 08, 2009 at 12:08:54AM +0200, Nicolas Sebrecht wrote:

quoted

IMHO, this improvement is broken by design. We try to fix a
receiver-only issue by a sender side fix.

I almost said the same thing: it is really the receiver's problem.
However, that doesn't mean the sender can't do simple things to help
hint the right thing to the receiver. For example, we already munge the
date fields to make sure the timestamp in each patch is increasing.

There's nothing simpler than slowing the rate of outgoing email. It doesn't
even have to be used; it is completely unintrusive and fully automated. It
even works with the confirmation. Best of all, it didn't take much code to
implement.

Sincerely,
Michael Witten

On Wed, Apr 8, 2009 at 09:25, Michael Witten [off-list ref] wrote:

For instance, with
a 14-patch series, one could set `--delay 60' and then let send-email
run happily for the next 14 minutes with nary a thought.

As implemented, it would only take 13 minutes.

For a 5-patch series:

* Patch 1/5 sent
* delay 60 seconds
* Patch 2/5 sent
* delay 60 seconds
* Patch 3/5 sent
* delay 60 seconds
* Patch 4/5 sent
* delay 60 seconds
* Patch 5/5 sent
* send email exits immediately

On Wed, Apr 08, 2009 at 09:25:25AM -0500, Michael Witten wrote:

Firstly, I presume that someone is electing to use this option, so it is
almost by definition not annoying for that person.

Sure, obviously only people who enable it will be affected. I was
thinking of it more in terms of group economics: how many people _will_
enable it, because they think the payoff outweighs the annoyance.

And perhaps you can argue that it has negligible group cost. The
code is small and not likely to introduce new bugs. So the only costs
are whatever any sender is willing to pay themselves. And in that sense,
if even one person uses it, it is a net win.

Secondly, it seems reasonable to drop into another VC, screen window, or
terminal instance, and then set send-email a-running. For instance, with
a 14-patch series, one could set `--delay 60' and then let send-email
run happily for the next 14 minutes with nary a thought.

Except for your 10-minute "no activity, so let's suspend to ram"
timeout. ;) But I think we can assume since the user is opting in to the
feature that they don't have a problem with it.

Frankly, I don't care how other people's patch series appear to me. I care
about how mine appear to others. Is this irrational? Probably, but I'm kind
of OC; I want my patch series to look like it's in order for everyone.

No, it's not irrational. We spend time proofreading, cleaning up
patches, etc, because we know other people will read them. And a little
effort by the sender can often save a lot of effort by many readers. And
people do it out of politeness, perhaps, or altruism, or simply because
it encourages others to spend the effort for them.

But I would personally find such a delay annoying (especially 60
seconds, which I consider to be huge), and I feel that by threading and
using sane date fields, I have done enough for the reader.

I have worked with 2 major mail clients that both display by date received:

  * Mac OS X's Mail
  * Gmail

Somebody reported in the thread I referenced earlier that gmane's web
interface has the same problem; I didn't check it myself, though.

There's nothing simpler than slowing the rate of outgoing email. It doesn't
even have to be used; it is completely unintrusive and fully automated. It
even works with the confirmation. Best of all, it didn't take much code to
implement.

I disagree that it's unintrusive (and I understand that you don't think
it is, and that there may be others like you, but I am pointing out
there are others who think the opposite). But given that it's optional,
and it's not very much code, I don't have a strong objection. My
original comment was that I was dubious whether it would work: however,
I was thinking you would set it to a few seconds. Setting it to 60
seconds, I can imagine it would have an impact.

-Peff

Jeff King [off-list ref] writes:

I disagree that it's unintrusive (and I understand that you don't think
it is, and that there may be others like you, but I am pointing out
there are others who think the opposite). But given that it's optional,
and it's not very much code, I don't have a strong objection. My
original comment was that I was dubious whether it would work: however,
I was thinking you would set it to a few seconds. Setting it to 60
seconds, I can imagine it would have an impact.

Heh, then why not make the option specify the number of _minutes_ not
seconds to delay?  That would help clarifying what this option is meant to
do.

The description for this new feature in the manual page should make it
absolutely clear that the maintainer who accepts such a feature is not
clueless to think that such a delay will guarantee something.  It is
merely for people who believe it would help guaranteeing the delivery
order _in practice_ and are willing to background the slow sending.

On Wed, Apr 08, 2009 at 09:25:25AM -0500, Michael Witten wrote:

This has nothing to do with what the receiver wants. This has everything
to do with what the sender wants. I want my patch series to be in order
even for wrongheaded receivers.

The --delay option may have an undesirable side effect. In case of
non-chained emails, unrelated mails could be insterted between patches
where *all* MUA would be affected. It's not only true for very high
volume message mailing-lists (million monkeys receiving...). FMPOV, it's
worse than all display issues we already know or have with the current
behaviour. 

-- 
Nicolas Sebrecht

On Thu, Apr 9, 2009 at 11:17, Nicolas Sebrecht
[off-list ref] wrote:

The --delay option may have an undesirable side effect. In case of
non-chained emails, unrelated mails could be insterted between patches
where *all* MUA would be affected. It's not only true for very high
volume message mailing-lists (million monkeys receiving...). FMPOV, it's
worse than all display issues we already know or have with the current
behaviour.

But it's already impossible to protect against this scenario. In that
situation, the smallest delay possible is desired, so --delay wouldn't
even be used (that is, its value would be zero). However, the transit
delay could never be small enough to guarantee that no other emails
are inserted into the patch series, so the only solution is to chain
them. At this point, we're back to the problem of arrival time, and
hence --delay becomes useful.

:-D

On Thu, Apr 9, 2009 at 12:27, Michael Witten [off-list ref] wrote:

On Thu, Apr 9, 2009 at 11:17, Nicolas Sebrecht
[off-list ref] wrote:

quoted

The --delay option may have an undesirable side effect. In case of
non-chained emails, unrelated mails could be insterted between patches
where *all* MUA would be affected. It's not only true for very high
volume message mailing-lists (million monkeys receiving...). FMPOV, it's
worse than all display issues we already know or have with the current
behaviour.

But it's already impossible to protect against this scenario. In that
situation, the smallest delay possible is desired, so --delay wouldn't
even be used (that is, its value would be zero). However, the transit
delay could never be small enough to guarantee that no other emails
are inserted into the patch series, so the only solution is to chain
them. At this point, we're back to the problem of arrival time, and
hence --delay becomes useful.

I do agree that --delay could exacerbate the spreading out of patches.

On Thu, Apr 09, 2009 at 12:27:18PM -0500, Michael Witten wrote:

On Thu, Apr 9, 2009 at 11:17, Nicolas Sebrecht
[off-list ref] wrote:

quoted

The --delay option may have an undesirable side effect. In case of
non-chained emails, unrelated mails could be insterted between patches
where *all* MUA would be affected. It's not only true for very high
volume message mailing-lists (million monkeys receiving...). FMPOV, it's
worse than all display issues we already know or have with the current
behaviour.

But it's already impossible to protect against this scenario.

But we could forbid the use of --delay with non-chained emails, no? Or
at least, warn it should not be used.

-- 
Nicolas Sebrecht

On Thu, 9 Apr 2009, Jeff King wrote:

On Wed, Apr 08, 2009 at 09:25:25AM -0500, Michael Witten wrote:

quoted

Firstly, I presume that someone is electing to use this option, so it is
almost by definition not annoying for that person.

Sure, obviously only people who enable it will be affected. I was
thinking of it more in terms of group economics: how many people _will_
enable it, because they think the payoff outweighs the annoyance.

My ISP doesn't allow me to send more than 20 emails at once.  So if a 
patch series has more than 20 patches, send-email will be denied the 
sending of the end of the series, which is rather anoying.

I didn't test this option (have no large patch series to post at the 
moment), but if it could allow me to easily work around the ISP 
restriction then I would be happy.


Nicolas

On Thu, 9 Apr 2009, Junio C Hamano wrote:

Jeff King [off-list ref] writes:

quoted

I disagree that it's unintrusive (and I understand that you don't think
it is, and that there may be others like you, but I am pointing out
there are others who think the opposite). But given that it's optional,
and it's not very much code, I don't have a strong objection. My
original comment was that I was dubious whether it would work: however,
I was thinking you would set it to a few seconds. Setting it to 60
seconds, I can imagine it would have an impact.

Heh, then why not make the option specify the number of _minutes_ not
seconds to delay?  That would help clarifying what this option is meant to
do.

Maybe only a few seconds is all that is needed in some circumstances.
For the issue I just posted, I wouldn't use it at all if the minimum 
delay was one minute.


Nicolas

On Thu, Apr 9, 2009 at 12:45, Nicolas Sebrecht
[off-list ref] wrote:

But we could forbid the use of --delay with non-chained emails, no? Or
at least, warn it should not be used.

Ah! Good point! I'll add that.