saving an image while serving

"But you cannot scale horizontally just simply"

Not working the way Nevin does. But that's not the Nevin's case. He's saying this "yes is one image, yes if it crashes we need to rollback, yes that architecture doesn't scale horizontally so what? is enough for the business it get us there just fine." (Nevin please correct me if I'm wrong)

Our case is different. Image persistence didn't fit our needs. We really needed to be able to scale horizontally. The way we found is to treat images like servers. Any user is be able to be attended by any image. That's how you do it.

Sticky sessions is a non issue once the httpd is configured (and properly balanced). About being forced to use a central data storage, that sounds plain wrong. I mean at design level. To be able to scale horizontally any time you have to design in a way that doesn't generate points of contention. How you design evading points of contention is an art. You need to find your most convenient way for your application. In our case we used one object database per group of people. Lots of groups, plenty of room for everybody's activity.

About memory size in images, we monitor the images and restart them as needed. The worst thing that can happen in an image restart is that you see your last request failed and in your next request, the http balancer sends you to another image. This is ACID so things gets done or do not. Nothing is lost.

That could happen with any image in any server (hardware).

About oversimplification and *most* scenarios: I'm sharing with you guys some details on how we got this thing done. What I'm not doing is trying to make people decide on "this" or "that" like some people suggested. Who the f* will invest or employ a guy deciding architecture like that? I'm making my case because is working and I think some could find it interesting or useful.

I thought it was clear but... hey, I'm counting with the intelligence and critical thinking of the reader. Always. 

Seriously, who decides the architecture of a startup in what he/she reads on a f* list?

Do your homework. If it f* works, then it does. If it doesn't, then it doesn't. 

On Apr 18, 2011, at 3:19 AM, Norbert Hartl wrote:

Am 17.04.2011 um 20:13 schrieb Sebastian Sastre:

The only realistic way we see for that is by scaling horizontally.

Many many worker images will do. And if people needs more, well... you simply add more.

I would really like to discuss this with Nevin and you. Most of the time you cannot "...simply add more" images. I would like to know what setups you have. In a web scenario where data is created by web clients in the image you can have it in one image, right. But you cannot scale horizontally just simply. You need to take special action like sticky sessions/domain partitioning if the domain model is not highly inter-connected or you forced to use a central data storage.
If you can solve it this way than you are lucky guys. Btw. is there a memory size in pharo from which performance will degrade faster?

I agree with you guys that you can achieve a lot with a simple image. But it sounded to me like an over-simplification of deployments which won't work in many scenarios.

Norbert

Of course that caution should be taken in the income/spending on validable* equations (monetization, customer adquisition, workers per server costs, etc) so things can grow smoothly.

*by validable equations I mean the math that comes out from the experience (and not math based on assumptions)

On Apr 17, 2011, at 6:29 AM, Norbert Hartl wrote:

Am 17.04.2011 um 01:57 schrieb Igor Stasenko:

This is really good to hear that.

I leaned something new. Usual way of design is to prevent failure(s).

But going that way we tend to forget that failures are inevitable, and

paying little attention towards what can be done

to restart/reload data from backups quickly and easily.

Indeed, embracing failure is the way we should design our systems.

Well, I can tell if you plan big systems than the failure scenarios planning takes a big part of the overall planning. There you embrace failures very very much. You can't go anywhere with something that is not redundant in at least one dimension. And the software has to fit in the hardware planning. So, not embracing failure is a pleasure that only hobbyists and smaller companies can have.

Norbert

On 17 April 2011 01:37, Nevin Pratt <[hidden email]> wrote:

Define "handful".  You'd be surprised.

We are a large enough company with enough orders, to support more than a

dozen employees, and a half a dozen contractors.  Our warehouse/office is

now about 26,000 square feet.  And, we're growing, and profitable, with 2011

looking to be our best year ever.

The exact numbers are confidential, but I am confident that we are the

largest company on the planet in our field (i.e., "Reborn Doll Supplies").

Besides, the orders aren't exactly "hand reentered".  I just didn't give

enough detail, because I didn't feel it was relevant.  We have an admin page

that can easily slurp the data in rapidly, if needed.  It's not hard to do.

And Squeak/Seaside has served us well.  And we just use the image for

persistence, as previously mentioned.  And we are very happy with it.

Nevin

Of course that only works when you have a handful of orders to reenter,

which isn't the case for most systems.

Sent from my iPhone

On 2011-04-16, at 12:41, "Sebastian Sastre" <[hidden email]>

wrote:

So, 15 minutes times 4 makes a whole hour per year.

Beautiful :)

Doesn't even justify to implement automation of the manual rollback (that

sounds feasible)

On Apr 16, 2011, at 1:29 PM, Nevin Pratt wrote:

Sebastion, that is *exactly* what I initially did at Bountiful Baby.

However, it has been several *years* since I've had to do your #4, so things

have changed slightly since.

Bountiful Baby is an eCommerce site, and the critical "object graph" (your

#2, below) information consists of inventory data (the website keeps track

of our inventory), and gift certificate data (for gift certificates that

have been issued).  Also, whenever either of those datums change, the

website sends an email-- for example, emails are (obviously) sent for each

order accepted, and, it sends emails when it issues a gift certificate.  So,

it is easy to discover (via the emails) what data was lost since the last

image crash.

Consequently, currently this is what happens:

1. image is saved from time to time (usually daily), and copied to a

separate "backup" machine.

2. if anything bad happens, the last image is grabbed, and the orders and/or

gift certificates that were issued since the last image save are simply

re-entered.

And, #2 has been *very* rarely done-- maybe a two or three times a year, and

then it turns out it is usually because I did something stupid.

For us, it's a whole lot easier to do persistence this way than bothering

with any persistence mechanism.  And, it turns out, it is *very* reliable,

with exactly one easily fixable glitch:

The glitch is: occasionally the production image UI will freeze, for no

known reason.  It doesn't effect Seaside, though, so the website keeps going

just fine.  And, if we run the "screenshot" app (in the configuration

screen), there is a link at the top for "Suspend UI Process", and "Resume UI

Process".  Just suspend and resume, and the UI becomes unstuck.

We've been doing persistence this way for years now, and I've been

*extremely* impressed with the reliability.

Before that, we used PostgreSQL and GLORP for persistence.  But I yanked

that code out years ago.  It wasn't worth the bother maintaining it.

If you have a daily image save, then on average there will be 12 hours of

lost data on a "random" crash.  Re-entering 12 hours of orders and/or gift

certificates (discovered from the emails, as mentioned above) might take 10

to 15 minutes.  Not a big deal at all for us.

Ten years ago I wouldn't have even considered doing persistence this way,

but I've changed.  Squeak has changed.  Seaside has changed.  And all for

the better.

Nevin

Given that you don't need transactions, to make that style work, I suggest

this:

1. save the image from time to time (as best suits your needs) like you're

doing now

2. use a secondary way to dump the object graph

3. use the normal image for as long as things are good

4. when shit happens you "transplant" the object graph into a new

"reincarnation" of your app in a fresh image

5. repeat

For dumping the ODB you have options: image segments, SIXX comes to mind now

If you make for yourself some "rescue" kit (script, tools, preloaded code in

fresh image), you can make 4 quite painless (or, why not, monitored and

automated)

sebastian

o/

On Apr 16, 2011, at 12:23 PM, Michal wrote:

hi -

Despite the warnings, I am really interested in sticking to the

simplest way of saving my seaside application data, ie periodically

saving and backuping the image. The seaside book states that

"saving [the image] while processing http requests is a risk you

want to avoid."

What is the status on that? Is that something we can fix? I have been

running an image in this mode for a few weeks, with no ill effect so

far, but I have had major problems with old image/vm combinations. So

is this something that might be fixed already?

Also, I recall that Avi had made a number of attempts at having an

image saved in a forked background process, eg

http://lists.squeakfoundation.org/pipermail/squeak-dev/2005-October/095547.html

did anybody pick up on this, or did anything come out of it?

thanks,

Michal

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

--

Best regards,

Igor Stasenko AKA sig.

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________
seaside mailing list
[hidden email]
http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________
seaside mailing list
[hidden email]
http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

Sebastian,

I personally was just fishing for deployment scenarios because I think it is important and could be added in some documentation. And that's why I was inquiring into depth. I think it is very valuable to know that you can just use image persistence if you meet the following criteria: Your model fits completely in the memory and concurrency is not that high. Or your model can be separated into distinct parts that you can spread over multiple images. Or you create some logging of the transcractions that you can pull back in if you experience a crash. Or you can rollback without much problems and that is feasible because crashes do not occurr that often.

I just didn't get Nevins approach. On the one hand he has everything in his image on the other hand he can "pull back in data" after experiencing a crash. That doesn't fit in my head. Nothing more I wanted to know.

Your statement that you can just simply scale is not proper without declaring the assumptions. And I would just want to hear your approach so it is documented as well and might enlight a lot of people like stubborn me.

Am 18.04.2011 um 15:59 schrieb Sebastian Sastre:

"But you cannot scale horizontally just simply"

Not working the way Nevin does. But that's not the Nevin's case. He's saying this "yes is one image, yes if it crashes we need to rollback, yes that architecture doesn't scale horizontally so what? is enough for the business it get us there just fine." (Nevin please correct me if I'm wrong)

It is not judging about architectural decisions. If it works it is fine. Who wants to argue?

Our case is different. Image persistence didn't fit our needs. We really needed to be able to scale horizontally. The way we found is to treat images like servers. Any user is be able to be attended by any image. That's how you do it.

Sticky sessions is a non issue once the httpd is configured (and properly balanced). About being forced to use a central data storage, that sounds plain wrong. I mean at design level. To be able to scale horizontally any time you have to design in a way that doesn't generate points of contention. How you design evading points of contention is an art. You need to find your most convenient way for your application. In our case we used one object database per group of people. Lots of groups, plenty of room for everybody's activity.

That's what I meant by saying domain partitioning. To me it doesn't matter if it is partioned in domains or access layers. If you think towards social networks your contention point free design might be hard to work out. 

About memory size in images, we monitor the images and restart them as needed. The worst thing that can happen in an image restart is that you see your last request failed and in your next request, the http balancer sends you to another image. This is ACID so things gets done or do not. Nothing is lost.

Well, that must be another approach than the "object database per group of people". 

That could happen with any image in any server (hardware).

About oversimplification and *most* scenarios: I'm sharing with you guys some details on how we got this thing done. What I'm not doing is trying to make people decide on "this" or "that" like some people suggested. Who the f* will invest or employ a guy deciding architecture like that? I'm making my case because is working and I think some could find it interesting or useful.

Again, I just asked because what you said wasn't clear to me. This confirms I'm someone that "find it interesting or useful". But in order to let me understand it you might need to answer some questions. If this is not what you want I can take it that way.

I thought it was clear but... hey, I'm counting with the intelligence and critical thinking of the reader. Always. 

You are right and the outcome of that is often a question.

Seriously, who decides the architecture of a startup in what he/she reads on a f* list?

I don't know. But I could imagine there are some people reading this list that are interested in exactly how you did it as an example of how things can be done. Most people will have a n-tier architecture as a background. It is hard to believe for them that you don't need it. And here you go and explain it. Than they have to decide if you are a genius or a jerk. That's up to them.

Do your homework. If it f* works, then it does. If it doesn't, then it doesn't. 

Agreed.

piece,

Norbert

On Apr 18, 2011, at 3:19 AM, Norbert Hartl wrote:

Am 17.04.2011 um 20:13 schrieb Sebastian Sastre:

The only realistic way we see for that is by scaling horizontally.

Many many worker images will do. And if people needs more, well... you simply add more.

I would really like to discuss this with Nevin and you. Most of the time you cannot "...simply add more" images. I would like to know what setups you have. In a web scenario where data is created by web clients in the image you can have it in one image, right. But you cannot scale horizontally just simply. You need to take special action like sticky sessions/domain partitioning if the domain model is not highly inter-connected or you forced to use a central data storage.
If you can solve it this way than you are lucky guys. Btw. is there a memory size in pharo from which performance will degrade faster?

I agree with you guys that you can achieve a lot with a simple image. But it sounded to me like an over-simplification of deployments which won't work in many scenarios.

Norbert

Of course that caution should be taken in the income/spending on validable* equations (monetization, customer adquisition, workers per server costs, etc) so things can grow smoothly.

*by validable equations I mean the math that comes out from the experience (and not math based on assumptions)

On Apr 17, 2011, at 6:29 AM, Norbert Hartl wrote:

Am 17.04.2011 um 01:57 schrieb Igor Stasenko:

This is really good to hear that.

I leaned something new. Usual way of design is to prevent failure(s).

But going that way we tend to forget that failures are inevitable, and

paying little attention towards what can be done

to restart/reload data from backups quickly and easily.

Indeed, embracing failure is the way we should design our systems.

Well, I can tell if you plan big systems than the failure scenarios planning takes a big part of the overall planning. There you embrace failures very very much. You can't go anywhere with something that is not redundant in at least one dimension. And the software has to fit in the hardware planning. So, not embracing failure is a pleasure that only hobbyists and smaller companies can have.

Norbert

On 17 April 2011 01:37, Nevin Pratt <[hidden email]> wrote:

Define "handful".  You'd be surprised.

We are a large enough company with enough orders, to support more than a

dozen employees, and a half a dozen contractors.  Our warehouse/office is

now about 26,000 square feet.  And, we're growing, and profitable, with 2011

looking to be our best year ever.

The exact numbers are confidential, but I am confident that we are the

largest company on the planet in our field (i.e., "Reborn Doll Supplies").

Besides, the orders aren't exactly "hand reentered".  I just didn't give

enough detail, because I didn't feel it was relevant.  We have an admin page

that can easily slurp the data in rapidly, if needed.  It's not hard to do.

And Squeak/Seaside has served us well.  And we just use the image for

persistence, as previously mentioned.  And we are very happy with it.

Nevin

Of course that only works when you have a handful of orders to reenter,

which isn't the case for most systems.

Sent from my iPhone

On 2011-04-16, at 12:41, "Sebastian Sastre" <[hidden email]>

wrote:

So, 15 minutes times 4 makes a whole hour per year.

Beautiful :)

Doesn't even justify to implement automation of the manual rollback (that

sounds feasible)

On Apr 16, 2011, at 1:29 PM, Nevin Pratt wrote:

Sebastion, that is *exactly* what I initially did at Bountiful Baby.

However, it has been several *years* since I've had to do your #4, so things

have changed slightly since.

Bountiful Baby is an eCommerce site, and the critical "object graph" (your

#2, below) information consists of inventory data (the website keeps track

of our inventory), and gift certificate data (for gift certificates that

have been issued).  Also, whenever either of those datums change, the

website sends an email-- for example, emails are (obviously) sent for each

order accepted, and, it sends emails when it issues a gift certificate.  So,

it is easy to discover (via the emails) what data was lost since the last

image crash.

Consequently, currently this is what happens:

1. image is saved from time to time (usually daily), and copied to a

separate "backup" machine.

2. if anything bad happens, the last image is grabbed, and the orders and/or

gift certificates that were issued since the last image save are simply

re-entered.

And, #2 has been *very* rarely done-- maybe a two or three times a year, and

then it turns out it is usually because I did something stupid.

For us, it's a whole lot easier to do persistence this way than bothering

with any persistence mechanism.  And, it turns out, it is *very* reliable,

with exactly one easily fixable glitch:

The glitch is: occasionally the production image UI will freeze, for no

known reason.  It doesn't effect Seaside, though, so the website keeps going

just fine.  And, if we run the "screenshot" app (in the configuration

screen), there is a link at the top for "Suspend UI Process", and "Resume UI

Process".  Just suspend and resume, and the UI becomes unstuck.

We've been doing persistence this way for years now, and I've been

*extremely* impressed with the reliability.

Before that, we used PostgreSQL and GLORP for persistence.  But I yanked

that code out years ago.  It wasn't worth the bother maintaining it.

If you have a daily image save, then on average there will be 12 hours of

lost data on a "random" crash.  Re-entering 12 hours of orders and/or gift

certificates (discovered from the emails, as mentioned above) might take 10

to 15 minutes.  Not a big deal at all for us.

Ten years ago I wouldn't have even considered doing persistence this way,

but I've changed.  Squeak has changed.  Seaside has changed.  And all for

the better.

Nevin

Given that you don't need transactions, to make that style work, I suggest

this:

1. save the image from time to time (as best suits your needs) like you're

doing now

2. use a secondary way to dump the object graph

3. use the normal image for as long as things are good

4. when shit happens you "transplant" the object graph into a new

"reincarnation" of your app in a fresh image

5. repeat

For dumping the ODB you have options: image segments, SIXX comes to mind now

If you make for yourself some "rescue" kit (script, tools, preloaded code in

fresh image), you can make 4 quite painless (or, why not, monitored and

automated)

sebastian

o/

On Apr 16, 2011, at 12:23 PM, Michal wrote:

hi -

Despite the warnings, I am really interested in sticking to the

simplest way of saving my seaside application data, ie periodically

saving and backuping the image. The seaside book states that

"saving [the image] while processing http requests is a risk you

want to avoid."

What is the status on that? Is that something we can fix? I have been

running an image in this mode for a few weeks, with no ill effect so

far, but I have had major problems with old image/vm combinations. So

is this something that might be fixed already?

Also, I recall that Avi had made a number of attempts at having an

image saved in a forked background process, eg

http://lists.squeakfoundation.org/pipermail/squeak-dev/2005-October/095547.html

did anybody pick up on this, or did anything come out of it?

thanks,

Michal

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

--

Best regards,

Igor Stasenko AKA sig.

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________

seaside mailing list

[hidden email]

http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________
seaside mailing list
[hidden email]
http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________
seaside mailing list
[hidden email]
http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside

_______________________________________________
seaside mailing list
[hidden email]
http://lists.squeakfoundation.org/cgi-bin/mailman/listinfo/seaside