Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
-
None
Description
When performing a restart of a large cluster where LDAP is being used
indirectly by nslcd, the LDAP servers are put under heavy load. This is more
evident in LDAP organizations that are large to begin with.
connection from pid=12345 uid=0 gid=0
nslcd_group_all()
myldap_search(base="cn=groups,cn=accounts,dc=corp,dc=local",
filter="(objectClass=posixGroup)")
ldap_result(): end of results
It turns out that these processes are the before-ANY hook script which runs when a service is started, like this one I was running locally to reproduce the query patterns.
/usr/bin/python2.6 /var/lib/ambari-agent/cache/stacks/HDP/2.0.6/hooks/before-
ANY/scripts/hook.py ANY /var/lib/ambari-agent/data/command-5950.json /var/lib
/ambari-agent/cache/stacks/HDP/2.0.6/hooks/before-ANY /var/lib/ambari-
agent/data/structured-out-5950.json INFO /var/lib/ambari-agent/data/tmp
I tracked the issue down to this function in resource_management/core/providers/accounts.py:
@property
def user_groups(self):
return [g.gr_name for g in grp.getgrall() if self.resource.username in g.gr_me
This property actually gets referenced at least 2 times for each user. The call to grp.getgrall() forces a complete enumeration of groups every time.
What this means is for a cluster with many nodes with many processes restarting across those nodes you are going to have many of these full enumeration searches running at the same time. In an enterprise with a large directory this will get very expensive, especially since this type of call is not cached by nscd.
I'm aware that the idiom used here to get the groups is common in python but it's actually pretty inefficient. Commands like id and groups have more efficient ways of discovering this. I'm not aware of the equivalent of these in Python.
@property
def user_groups(self):
ret = []
(rc, output) = shell.checked_call(['groups', self.resource.username](https://h
sudo=True)
if rc == 0:
ret.extend(output.split(':')[1](
).lstrip().split())
return ret
This converts the full LDAP scan for groups to more efficient queries targeted
to the user. The lookups done by the groups command are also 100% cacheable.
Since it's a checked call the `rc == 0` check is probably not needed.
An unfortunate effect of how usermod and friends work is that it always
invalidates the nscd cache after it's run. This means that Ambari could still
be a lot more efficient than it is when LDAP is in play by being pickier about
when it runs commands like useradd/usermod/groupadd/groupmod.
We can also probably put a timed cache on the results from `grp.getgrall()` or
`groups` in memory, configurable by the agent config file. This way, we would
only call it once every hour or so.
Attachments
Attachments
Issue Links
- links to
