lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Tomás Fernández Löbbe (JIRA) <j...@apache.org>
Subject [jira] [Commented] (SOLR-13439) Make collection properties easier and safer to use in code
Date Thu, 09 May 2019 04:56:00 GMT

    [ https://issues.apache.org/jira/browse/SOLR-13439?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16836080#comment-16836080
] 

Tomás Fernández Löbbe commented on SOLR-13439:
----------------------------------------------

None of those cases use the feature as it's supposed to be used. You either need frequent
access to the properties (i.e. per request, timer, something) or not (i.e. maybe to validate
something after a particular type of request, like a collection API with a particular set
of values, etc)

Case 1 (Frequent access):
 1. T+0: SolrCore starts (i.e. A replica is added, a collection is created, you name it)
 2. T+0: On initialization, a component that relies on collection properties registers a listener.
Solr reads from ZooKeeper once.
 3. T+0 to T+end-of-life-of-the-core: The watch remains.
 4. T+end-of-life-of-the-core: no more reads are expected, the watch is removed.
 Times read with current approach: 1 + Number of modifications of properties.
 Times with the cache approach: unknown, at least the same as with the listener, but depends
on when the properties are accessed.

Case 2: Don't set a watch. (you only do this if you know you are going to access the properties
a low number of times, so low that it doesn't make sense to spend resources in watching).
 T + 0: read the properties (reads from Zookeeper)
 done.
{quote}Code that calls getCollectionProperties() is either fast or slow depending on whether
or not someone else has set a watch on that collection.
{quote}
If your component doesn't set a listener, assume it's going to ZooKeeper. You'd only do this
if reads to a property are infrequent enough (see case 2 above)
{quote}I think the current inconsistency is worse because something that was fast due to the
action of unrelated code (getCollectionProperties()) can become surprisingly slow,
{quote}
Again, assume {{getCollectionProperties()}} goes to ZooKeeper, but this is only used in case
of infrequent access. The caller makes the informed tradeoff of going to ZooKeeper per call
instead of spending resources (ZooKeeper’s and Solr’s) in watching.
{quote}whereas after my patch, setting the watch may become surprisingly fast due to the effect
of unrelated code.
{quote}
... if it's in cache you mean?
{quote}in no case will it be more
{quote}
I can give you two cases:
 Case 1: Collection properties are accessed infrequently (like in my “case 2 above”),
but collection properties change frequently (i.e. every second)
 1. T + 0: call to getCollectionProperties(), Zk watch is set and element is on cache
 2. T + 1 to T + 9: Collection properties changes, fires watches to Solr. Solr receives the
watch and reads from Zookeeper
 3. T + 10 cache expires
 With cache, we read from Zookeeper 10 times, and ZooKeeper fires 10 watches. Without cache,
we read once, ZooKeeper doens't fire any watch. Keep in mind that some clusters may have many
collections (hundreds/thousands?), this may add a lot of load to ZooKeeper for things that
aren’t going to be needed. 
 Case 2: A component doesn’t rely on a listener, but relies on cache. 
 1. T + 0: call to getCollectionProperties(), Zk watch is set and element is on cache
 2. T + 10, cache expires
 3. T + 11: call to getCollectionProperties(), Zk watch is set and element is on cache
 4. T + 20, cache expires
 5. …
 With a listener, this is just one read. With cache, this is, again, unknown, but up to N,
the number of calls to {{getCollectionProperties()}}

> Make collection properties easier and safer to use in code
> ----------------------------------------------------------
>
>                 Key: SOLR-13439
>                 URL: https://issues.apache.org/jira/browse/SOLR-13439
>             Project: Solr
>          Issue Type: Improvement
>      Security Level: Public(Default Security Level. Issues are Public) 
>          Components: SolrCloud
>    Affects Versions: master (9.0)
>            Reporter: Gus Heck
>            Assignee: Gus Heck
>            Priority: Major
>         Attachments: SOLR-13439.patch, SOLR-13439.patch
>
>
> (breaking this out from SOLR-13420, please read there for further background)
> Before this patch the api is quite confusing (IMHO):
>  # any code that wanted to know what the properties for a collection are could call zkStateReader.getCollectionProperties(collection)
but this was a dangerous and trappy API because that was a query to zookeeper every time.
If a naive user auto-completed that in their IDE without investigating, heavy use of zookeeper
would ensue.
>  # To "do it right" for any code that might get called on a per-doc or per request basis
one had to cause caching by registering a watcher. At which point the getCollectionProperties(collection) magically
becomes safe to use, but the watcher pattern probably looks famillar induces a user who hasn't
read the solr code closely to create their own cache and update it when their watcher is
notified. If the caching side effect of watches isn't understood this will lead to many in-memory
copies of collection properties maintained in user code.
>  # This also creates a task to be scheduled on a thread (PropsNotification) and induces
an extra thread-scheduling lag before the changes can be observed by user code.
>  # The code that cares about collection properties needs to have a lifecycle tied to
either a collection or something other object with an even more ephemeral life cycle such
as an URP. The user now also has to remember to ensure the watch is unregistered, or there
is a leak.
> After this patch
>  # Calls to getCollectionProperties(collection) are always safe to use in any code anywhere.
Caching and cleanup are automatic.
>  # Code that really actually wants to know if a collection property changes so it can
wake up and do something (autoscaling?) still has the option of registering a watcher that
will asynchronously send them a notification.
>  # Updates can be observed sooner via getCollectionProperties with no need to wait for
a thread to run. (vs a cache held in user code)



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org


Mime
View raw message