Forum OpenACS Q&A: Lazy site node caching

Sadly I am not happy with the solution as it requires a trip to the database. Why ?

The method I use is to go from the root node and for each subsequent node in the URL I will load the subpart into the database if it is not already there. I stop the moment I am not finding the next part in the database and return with the previous node. As you can see this results in any case in a trip to the database, unless we are querying the root of a package. Why is that? The site node cache is only for packages and folders, it does not take into account the file names. So if you have /file-storage/folder-view, /file-storage would be in the site_nodes table, folder-view would not. But there is no way for me to predetermine if it is or not unless I hit the database.

The next thought I have now is to add a "no_children_p" entry to the site_nodes NSV. This will tell me that there are no more children for the particular site_node I am looking at so I do not have to do the last look at the database. As the site node is flushed whenever I add a child, I should be save from making a lot of code changes just to keep track of the variable.

Does this approach make sense to you? Did I miss something? And does this affect the XoTCL request processor, do I need to make changes to xotcl-core as well, as I am changing site_node::get_from_url as the only place to query the site_nodes NSV at the moment.

We are very interested on your work.
One of the problems with our cluster
(well, THE PROBLEM I said) is with site node
cache.

If you want we test some code, please tell us.

Thanks,
Agustin

We are testing it and I get the first problem. The function
dotlrn::is_package_mounted is not detecting the attach mounted below dotlrn package in the OpenACS start process.
Next is the error log:

[13/Mar/2007:12:56:41][14966.16384][-main-] Notice: Loading packages/dotlrn/tcl/dotlrn-init.tcl...
[13/Mar/2007:12:56:41][14966.16384][-main-] Notice: dotlrn-init: starting...
[13/Mar/2007:12:56:41][14966.16384][-main-] Notice: dotlrn-init: attachments being automounted at /dotlrn/attach
[13/Mar/2007:12:56:41][14966.16384][-main-] Notice: dotlrn::mount_package: object_type apm_package url /dotlrn/ object_id 2121 instance_name dotLRN
package_type apm_application package_id 2121 name dotlrn node_id 2120 has_children_p 1 directory_p t package_key dotlrn pattern_p t parent_id 498
NOTICE: adding missing FROM-clause entry for table "acs_object_id_seq"
CONTEXT: PL/pgSQL function "acs_object__new" line 17 at SQL statement
PL/pgSQL function "site_node__new" line 23 at assignment
[13/Mar/2007:12:56:41][14966.16384][-main-] Error: Ns_PgExec: result status: 7 message: ERROR: duplicate key violates unique constraint "site_nodes_un"
CONTEXT: SQL statement "INSERT INTO site_nodes (node_id, parent_id, name, object_id, directory_p, pattern_p) values ( $1 , $2 , $3 , $4 , $5 , $6 )"
PL/pgSQL function "site_node__new" line 35 at SQL statement

[13/Mar/2007:12:56:41][14966.16384][-main-] Error: dbinit: error(pizarradb.uv.es:5433:openacsdb_5_2_desa,ERROR: duplicate key violates unique constraint
"site_nodes_un"
CONTEXT: SQL statement "INSERT INTO site_nodes (node_id, parent_id, name, object_id, directory_p, pattern_p) values ( $1 , $2 , $3 , $4 , $5 , $6 )"
PL/pgSQL function "site_node__new" line 35 at SQL statement
): '

select site_node__new(NULL,'2120','attach',NULL,'t','t',NULL,NULL)
...

Regards,
Agustín

if {[catch {nsv_get site_nodes "${new_url}/"} result] == 0} {
set node_id ""
} else {
if {$new_node(has_children_p) && [lsearch $acs_subsite_dir_list $name] == -1} {
set node_id [db_string node_id "select node_id from site_nodes where parent_id = :parent_id and name=:name" -default ""]
ns_log Debug "Loading from the database $test_url $name $parent_id"
} else {
set node_id ""
}
}

Your code works. At this moment all works except:

-Some portles displaying the msg (example):
Error in include template "/var/lib/aolserver/oacs_5_2/packages/lorsm/lib/user-lorsm": can't read "url_by_node_id(17673735)": no such element in array

-xotcl:
site_node::get "must pass in either url or node_id"
while executing
"error "site_node::get \"must pass in either url or node_id\"""
(procedure "site_node::get" line 4)
invoked from within
"site_node::get -url $mount_url"
(procedure "::Generic::package_id_from_package_key" line 4)
invoked from within
"::Generic::package_id_from_package_key xotcl-request-monitor"

I will like resolve these problems before test the performance in our true cluster.

Any pointer?
Agustín

We are working in our production cluster with your code
with some small mods. The speed are very increased.
Thanks!

We have since 400 concurrent connections.
If anybody want to know how is our system:

http://aulavirtual.uv.es/ficheros/view/imagenes%5C/CLUSTER_AulaVirtual_pub.gif

NOTES:

Another patch (line 360, site-nodes-procs.tcl):
db_foreach $query_name {} {
if {$parent_id eq ""} {
# url of root node
set url "/"
} else {
# append directory to url of parent node
if { [info exists url_by_node_id($parent_id)] } {
set url $url_by_node_id($parent_id)
} else {
set url [db_string snid {select site_node__url(:parent_id)} -default ""]
}
append url $name
if { $directory_p eq "t" } { append url "/" }
}

Regards,
Agustin

.LRN Honchos. Is it okay to include the patch above in 5.3 or should I only commit this to the HEAD version of .LRN?

The main problem here is that from the URL you cannot detect if you are looking at a site_node or a folder within a package.

So, if you go from top to bottom, I already implemented a check for "has_children_p". Sadly this does not help if you work with multiple subsites e.g /malte. /malte/register would look at /malte, see it has children (e.g. /malte/photo), therefore try to load /malte/register from the database, which will obviously fail because /register is a directory, not a site_node.

Next idea is to try and load the children of a site_node not mounted under / and /dotlrn/, /dotlrn/clubs, /dotlrn/communities (you get the picture). Then I could say "children_loaded_p" and if I come from the top and hit a "children_loaded_p" in the site_node cache then I assume that all children have been loaded and go the usual way of finding a site_node, which assumes that it is in the site_node cache.

Only if this fails I would try to reload the site_node cache for the last node (from bottom to top) that says "children_loaded_p". This is due to the reason that you might want to mount a new package under an already loaded site_node, which would then not be found by the cluster.

ANy other idea is highly welcome though, but I am out of ideas for how to make good assumption if I need to hit the database for a URL like /malte/forums/moderate/edit or not.

We are testing a bit your code.
Only we have detected some problems:

-The images in
http://www.xx.xx/dotlrn/configure
for up, down, delete, ... are not shown
until we refresh the cache with
http://www.xx.xx/acs-admin/cache

Same behaviour with urls
portal/admin/portal-config?portal_id=xxx
below
http://www.xx.xx/dotlrn/admin/portal-templates

-We are geting the next messages in the log:
[30/May/2007:09:38:51][23199.1106975072][-conn:20-] Notice: Loading from the database /dotlrn/classes/c032/12254/c07c032a12254gH c07c032a12254gH 4890528
[30/May/2007:09:38:54][23199.1106975072][-conn:20-] Error: Tcl exception:
can't read "node(url)": no such element in array
while executing
"string equal $node(url) "[ad_conn url]/""
(procedure "rp_filter" line 92)
invoked from within
"rp_filter preauth"
...

Regards,
Agustin

I suspect this problem comes from the site_nodes library that both dotlrn and new-portal implement (yes! it's duplicated in those 2 packages...). This should be removed and the calls to those procs replaced by calls to site-node ones (acs-tcl).

I've planned to work on this some time during the next 2 weeks. Not sure I will have time though.

I've finished the cleanup of site_nodes and committed on HEAD. The duplicated libraries have been removed and calls to site_nodes::* replaced by calls to site_node::

I just did some smoke tests: dotlrn install, create communities, add applets, create users and add them to communities, etc. (where those calls were used) and everything seem to work correctly.

Did you reinstall first? It won't fix what you describe but the dotlrn installation was broken because of site_nodes.

Anyway, I've noticed this behavior too. I was about to post about it after doing more tests.

On my side, I get a "file not found" when I go to the admin pages, but I don't need to restart aolserver, just to reload the page (I'm using firefox 2 on debian). It happens also on openacs when applications are added, so I guess there are still issues with caching.

Malte, any ideas?

After re-installing the db using the oacs-5-3 version of site-nodes-procs* (to rule out any residual side effects), I see this:

/api-doc/ file not found
/api-doc/index file not found
/api-doc/index? the usual page shows up.

but that probably doesn't help much for test cases with the current code.

Am trying again with current cvs, since there appears to be some recent changes in those files..

By current version, do you mean 5.3?

I have the "file not found" problem but with HEAD code, not with 5.3 one.