lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Erick Erickson (JIRA)" <j...@apache.org>
Subject [jira] [Updated] (SOLR-6512) Add a collections API call to attach arbitrary properties to a replica
Date Sat, 13 Sep 2014 21:39:33 GMT

     [ https://issues.apache.org/jira/browse/SOLR-6512?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Erick Erickson updated SOLR-6512:
---------------------------------
    Description: 
This is a sub-task for SOLR-6491, but seems generally useful. 

Since this is in support of the "preferredLeader" functionality, I've run into some considerations
that I wanted some feedback on how to handle.

"preferredLeader" has the restriction that there should only be one per slice, so setting
this for a particular node means removing the property for all the other replicas on the slice.
Not a problem to do, my question is more whether this is something reasonable to enforce on
an arbitrary property based on what that property is? Perfectly do-able, but "semantically
challenged". Currently, this is never a node with "preferedLeader" set to "false", it is forcibly
removed from other nodes in the slice when this property is assigned.

The problem here is that there's nothing about assigning an arbitrary property to a node that
would reasonably imply this kind of behavior. One could always control this with secondary
flags on the command, e.g. "shardExclusive=true|false" for instance, perhaps with safety checks
in for known one-per-shard properties like "preferredLeader".

"preferredLeader" seems to fit more naturally into a "role", but currently ADDROLE and DELTEROLE
have nothing to do with the notion of setting a role for a particular node relative to a collection/shard.
Easy enough to add, but enforcing the "only one node per slice may have this role" rule there
is similarly arbitrary and overloads the ADDROLE/DELETEROLE in a way that seems equally confusing.
Plus, checking whether the required collection/shard/node params are present becomes based
on the value of the property being set, which is all equally arbitrary.

The other interesting thing is that setting an arbitrary property on a node would allow one
to mess things up royally by, say, changing properties like "core", or "base_url" or node_name
at will. Actually this is potentially useful, but very, very dangerous and I'm not particularly
interested in supporting it ;).  I suppose we could require a prefix, say the only settable
properties are "property.whatever".

We could also add something specific to nodes, something like ADDREPLICAROLE/DELETEREPLICAROLE,
perhaps with sub-params like "onlyOneAllowedPerShard", but this gets messy and relies on the
users "doing the right thing". I prefer enforcing rules like this  based on the role I think.
Or at least enforcing these kinds of requirements on the "preferredLeader" role if we go that
way.

What are people's thoughts here? I think I'm tending towards the ADDREPLICAROLE/DELETEREPLICAROLE
way of doing this, but it's not set in stone. I have code locally for arbitrary properties
that I can modify for the role bits.

So, if I'm going to summarize the points I'd like feedback on:
1> Is setting arbitrary properties on a node desirable? If so, should we require a prefix
like "property" to prevent resetting values SolrCloud depends on?

2> Is it better to piggyback on ADDROLE/DELETEROLE? Personally I'm not in favor of this
one. Too messy with requiring additional parameters to work right in this case

3> Is the best option to create new collections API calls for ADDREPLICAROLE/DELETEREPLICAROLE
that
3.1> require collection/slice/node parameters
3.2> enforces the "onlyOnePerShard" rule for certain known roles
3.3 v1> allows users to specify arbitrary roles something like "onlyOnePerShard" as an
optional T|F parameter, otherwise is totally open.
-or-
3.3 v2> No support other than "preferredLeader", only roles that are pre-defined are allowed,
in which case the "onlyOnePerShard" is implicit in the role.






  was:
This is a sub-task for SOLR-6491, but seems generally useful. 

Since this is in support of the "preferredLeader" functionality, I've run into some considerations
that I wanted some feedback on how to handle.

"preferredLeader" has the restriction that there should only be one per slice, so setting
this for a particular node means removing the property for all the other replicas on the slice.
Not a problem to do, my question is more whether this is something reasonable to enforce on
an arbitrary property based on what that property is? Perfectly do-able, but "semantically
challenged". Currently, this is never a node with "preferedLeader" set to "false", it is forcibly
removed from other nodes in the slice when this property is assigned.

The problem here is that there's nothing about assigning an arbitrary property to a node that
would reasonably imply this kind of behavior. One could always control this with secondary
flags on the command, e.g. "shardExclusive=true|false" for instance, perhaps with safety checks
in for known one-per-shard properties like "preferredLeader".

"preferredLeader" seems to fit more naturally into a "role", but currently ADDROLE and DELTEROLE
have nothing to do with the notion of setting a role for a particular node relative to a collection/shard.
Easy enough to add, but enforcing the "only one node per slice may have this role" rule there
is similarly arbitrary and overloads the ADDROLE/DELETEROLE in a way that seems equally confusing.
Plus, checking whether the required collection/shard/node params are present becomes based
on the value of the property being set, which is all equally arbitrary.

The other interesting thing is that setting an arbitrary property on a node would allow one
to mess things up royally by, say, changing properties like "core", or "base_url" or node_name
at will. Actually this is potentially useful, but very, very dangerous and I'm not particularly
interested in supporting it ;).  I suppose we could require a prefix, say the only settable
properties are "property.whatever".

We could also add something specific to nodes, something like ADDREPLICAROLE/DELETEREPLICAROLE,
perhaps with sub-params like "onlyOneAllowedPerShard", but this gets messy and relies on the
users "doing the right thing". I prefer enforcing rules like this  based on the role I think.
Or at least enforcing these kinds of requirements on the "preferredLeader" role if we go that
way.

What are people's thoughts here? I think I'm tending towards the ADDREPLICAROLE/DELETEREPLICAROLE
way of doing this, but it's not set in stone. I have code locally for arbitrary properties
that I can modify for the role bits.

So, if I'm going to summarize the points I'd like feedback on:
1> Is setting arbitrary properties on a node desirable? If so, should we require a prefix
like "property" to prevent resetting values SolrCloud depends on?

2> Is it better to piggyback on ADDROLE/DELETEROLE? Personally I'm not in favor of this
one. Too messy with requiring additional parameters to work right in this case

3> Is the best option to create new collections API calls for ADDREPLICAROLE/DELETEREPLICAROLE
that
3.1> require collection/slice/node parameters
3.2> enforces the "onlyOnePerShard" rule for certain known roles
3.3> allows users to specify something like "onlyOnePerShard" as an optional T|F parameter





> Add a collections API call to attach arbitrary properties to a replica
> ----------------------------------------------------------------------
>
>                 Key: SOLR-6512
>                 URL: https://issues.apache.org/jira/browse/SOLR-6512
>             Project: Solr
>          Issue Type: Improvement
>            Reporter: Erick Erickson
>            Assignee: Erick Erickson
>
> This is a sub-task for SOLR-6491, but seems generally useful. 
> Since this is in support of the "preferredLeader" functionality, I've run into some considerations
that I wanted some feedback on how to handle.
> "preferredLeader" has the restriction that there should only be one per slice, so setting
this for a particular node means removing the property for all the other replicas on the slice.
Not a problem to do, my question is more whether this is something reasonable to enforce on
an arbitrary property based on what that property is? Perfectly do-able, but "semantically
challenged". Currently, this is never a node with "preferedLeader" set to "false", it is forcibly
removed from other nodes in the slice when this property is assigned.
> The problem here is that there's nothing about assigning an arbitrary property to a node
that would reasonably imply this kind of behavior. One could always control this with secondary
flags on the command, e.g. "shardExclusive=true|false" for instance, perhaps with safety checks
in for known one-per-shard properties like "preferredLeader".
> "preferredLeader" seems to fit more naturally into a "role", but currently ADDROLE and
DELTEROLE have nothing to do with the notion of setting a role for a particular node relative
to a collection/shard. Easy enough to add, but enforcing the "only one node per slice may
have this role" rule there is similarly arbitrary and overloads the ADDROLE/DELETEROLE in
a way that seems equally confusing. Plus, checking whether the required collection/shard/node
params are present becomes based on the value of the property being set, which is all equally
arbitrary.
> The other interesting thing is that setting an arbitrary property on a node would allow
one to mess things up royally by, say, changing properties like "core", or "base_url" or node_name
at will. Actually this is potentially useful, but very, very dangerous and I'm not particularly
interested in supporting it ;).  I suppose we could require a prefix, say the only settable
properties are "property.whatever".
> We could also add something specific to nodes, something like ADDREPLICAROLE/DELETEREPLICAROLE,
perhaps with sub-params like "onlyOneAllowedPerShard", but this gets messy and relies on the
users "doing the right thing". I prefer enforcing rules like this  based on the role I think.
Or at least enforcing these kinds of requirements on the "preferredLeader" role if we go that
way.
> What are people's thoughts here? I think I'm tending towards the ADDREPLICAROLE/DELETEREPLICAROLE
way of doing this, but it's not set in stone. I have code locally for arbitrary properties
that I can modify for the role bits.
> So, if I'm going to summarize the points I'd like feedback on:
> 1> Is setting arbitrary properties on a node desirable? If so, should we require a
prefix like "property" to prevent resetting values SolrCloud depends on?
> 2> Is it better to piggyback on ADDROLE/DELETEROLE? Personally I'm not in favor of
this one. Too messy with requiring additional parameters to work right in this case
> 3> Is the best option to create new collections API calls for ADDREPLICAROLE/DELETEREPLICAROLE
that
> 3.1> require collection/slice/node parameters
> 3.2> enforces the "onlyOnePerShard" rule for certain known roles
> 3.3 v1> allows users to specify arbitrary roles something like "onlyOnePerShard" as
an optional T|F parameter, otherwise is totally open.
> -or-
> 3.3 v2> No support other than "preferredLeader", only roles that are pre-defined are
allowed, in which case the "onlyOnePerShard" is implicit in the role.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

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


Mime
View raw message