{toc}
h1. Executive Summary
*How changing a cfloop from a query loop to an index loop eliminated this application's memory problem*
A process in one of our internal tools takes a while to run. In this case, a specific instance was crashing due to Java Heap Space errors. The image on the top shows the heap \-\- with a snapshot taken while the application was in its death throes \-\- prior to fixing the problem. The image on the bottom shows the heap \-\- with a snapshot taken at roughly the same time during the application's run \-\- after making 2 small code changes. One server is dead; the other is alive. This is our story.
!deadserver_1.png!
h1. The code, before
This particular process works in two parts that both do the same kind of thing
# Query another database and fetch around 350k rows. Loop over those rows and "do stuff"
# Query another database and fetch around 950k rows. Loop over those rows and "do other stuff"
The app was crashing in that second part, right around 770k rows. Here's a snippet:
{code:title=The Killing Code}<cfloop query="records" startrow="#startRow#" endrow="#endRow#">
<cfset recCount++ />
<cfset row = doStuff(records, records.currentrow, recCount)>
<cfset arrayAppend( rows, row )>
</cfloop>
{code}It's your standard cfloop over a query.
h1. Heap memory, before
Here's what it looked like in Eclipse MAT
As the app was dying, I took a heap dump (*[described here... it's easy|http://blog.mxunit.org/2010/01/using-eclipse-mat-to-track-down.html]*) and saw that a thing called coldfusion.runtime.CFDummyComponent was consuming over 800MB of Ram. Drilling down a few levels, I saw thousands of "coldfusion.tagext.lang.LoopTag". Drilling into one of those, I saw that for each loop tag, there were 5000 coldfusion.sql.imq.Row objects. 5000 because that's how many records we were looping over in that cfloop.
!deadserver_looptags.png!
h1. The code, after
After seeing that all of the memory was being consumed inside of loop tags, I looked at the two cfloops that were running. My thought process was something like, "If we're doing a cfloop query='', and if the memory analyzer is telling me that each loop is retaining thousands of objects related to that query, then how can I perform the same loop without giving the query to the loop?" It should come as no surprise that the answer was:
{code:language=cf|title=PartnerMigrationCommand.cfc|controls=true|linenumbers=true}<cfloop from="#startRow#" to="#endRow#" index="i">
<cfset recCount++ />
<cfset row = doStuff(records, i, recCount)>
<cfset arrayAppend( rows, row )>
</cfloop>
{code}
h1. Heap memory, after
As proof that this code change had the intended effect, I ran the app again and took a heap snapshot at roughly the same point. The results were not surprising... not a single "looptag" in sight.
What's more, during the running of the application, *retained heap dropped from 882 MB to 52 MB*, simply by changing the cfloop style.
!deadserver_No_looptags.png!
h1. Conclusion
I'm not suggesting that you should go change your loops. Here's what I want you to takeaway:
# When debugging, don't go flailing about, changing code willy nilly. *Get proof*
# Try to think creatively about the information you see. In this case, I saw a bunch of loop tags, leading me to ask, "what could be going on with these loops?". Assume nothing
# Make small changes. Change one thing at a time, run the app, take measurements, and see their effects
I do believe this is a bug. If you agree, [please vote|http://cfbugs.adobe.com/cfbugreport/flexbugui/cfbugtracker/main.html#bugId=86444] .
h1. Executive Summary
*How changing a cfloop from a query loop to an index loop eliminated this application's memory problem*
A process in one of our internal tools takes a while to run. In this case, a specific instance was crashing due to Java Heap Space errors. The image on the top shows the heap \-\- with a snapshot taken while the application was in its death throes \-\- prior to fixing the problem. The image on the bottom shows the heap \-\- with a snapshot taken at roughly the same time during the application's run \-\- after making 2 small code changes. One server is dead; the other is alive. This is our story.
!deadserver_1.png!
h1. The code, before
This particular process works in two parts that both do the same kind of thing
# Query another database and fetch around 350k rows. Loop over those rows and "do stuff"
# Query another database and fetch around 950k rows. Loop over those rows and "do other stuff"
The app was crashing in that second part, right around 770k rows. Here's a snippet:
{code:title=The Killing Code}<cfloop query="records" startrow="#startRow#" endrow="#endRow#">
<cfset recCount++ />
<cfset row = doStuff(records, records.currentrow, recCount)>
<cfset arrayAppend( rows, row )>
</cfloop>
{code}It's your standard cfloop over a query.
h1. Heap memory, before
Here's what it looked like in Eclipse MAT
As the app was dying, I took a heap dump (*[described here... it's easy|http://blog.mxunit.org/2010/01/using-eclipse-mat-to-track-down.html]*) and saw that a thing called coldfusion.runtime.CFDummyComponent was consuming over 800MB of Ram. Drilling down a few levels, I saw thousands of "coldfusion.tagext.lang.LoopTag". Drilling into one of those, I saw that for each loop tag, there were 5000 coldfusion.sql.imq.Row objects. 5000 because that's how many records we were looping over in that cfloop.
!deadserver_looptags.png!
h1. The code, after
After seeing that all of the memory was being consumed inside of loop tags, I looked at the two cfloops that were running. My thought process was something like, "If we're doing a cfloop query='', and if the memory analyzer is telling me that each loop is retaining thousands of objects related to that query, then how can I perform the same loop without giving the query to the loop?" It should come as no surprise that the answer was:
{code:language=cf|title=PartnerMigrationCommand.cfc|controls=true|linenumbers=true}<cfloop from="#startRow#" to="#endRow#" index="i">
<cfset recCount++ />
<cfset row = doStuff(records, i, recCount)>
<cfset arrayAppend( rows, row )>
</cfloop>
{code}
h1. Heap memory, after
As proof that this code change had the intended effect, I ran the app again and took a heap snapshot at roughly the same point. The results were not surprising... not a single "looptag" in sight.
What's more, during the running of the application, *retained heap dropped from 882 MB to 52 MB*, simply by changing the cfloop style.
!deadserver_No_looptags.png!
h1. Conclusion
I'm not suggesting that you should go change your loops. Here's what I want you to takeaway:
# When debugging, don't go flailing about, changing code willy nilly. *Get proof*
# Try to think creatively about the information you see. In this case, I saw a bunch of loop tags, leading me to ask, "what could be going on with these loops?". Assume nothing
# Make small changes. Change one thing at a time, run the app, take measurements, and see their effects
I do believe this is a bug. If you agree, [please vote|http://cfbugs.adobe.com/cfbugreport/flexbugui/cfbugtracker/main.html#bugId=86444] .