[
https://issues.apache.org/jira/browse/HBASE-9291?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13860731#comment-13860731
]
Andrew Purtell commented on HBASE-9291:
---------------------------------------
First let me clarify my above second suggestion: We could hang a map accessible
to all CPs in the RS off of RegionServerServices as like was done for the
region level in HBASE-6505. Then the client would provide the (large) state as
an attribute on only the first mutation sent to each regionserver. (More on
this below.) The CP would observe the attribute and apply it to RS-level shared
state. Then the mutation and subsequent mutations could be processed referring
to the updated RS-level state.
The client side is tricky.
bq. We could try to figure out which Put is the "first one" for each region,
but what if a split occurs after we figure this out – this seems too brittle.
If we introduce a new client API to the effect of "send one RPC to each RS",
then this amounts to a modified coprocessor endpoint execution, but with an
invocation target that is a singleton to each RS, and should be subject to the
same security considerations. Passing an attribute on the first put to a RS
sidesteps the need for EXEC grants (HBASE-6104) on any endpoint invocation
target, which is what sounds like the goal you are after.
Whether an endpoint invocation or a mutation, we have the same issue that the
local knowledge of cluster state can at any point be stale. Live servers can
come and go, and regions can move around, and there is no transactional state
update protocol running between clients and servers for updating this
information. Even if there were, cluster topology can change mid flight. A
"send one RPC to each RS" API could miss a newly onlined server that came up
after the call(s) started and yet opened some relevant regions asynchronously.
Whether trying to figure out which put is the first for a RS, or selecting keys
for a set of coprocessor endpoints such that you only invoke one per RS, or
using a new "send one RPC to each RS", on the server you'd have to handle the
same set of issues, right? There could be 0, 1, or ~2 large data transfers per
RS:
- 0 if a new server is onlined and regions are assigned after the put or "send
one RPC to each RS" calls are in progress
- 1 if the cluster topology is unchanged over the entire client action
- ~2 if a region is moved or split, or even in the case of one-RPC-per-server
if there is a RPC retry on account of the failed transmission back to the
client of a server side success indication
I wouldn't use the word brittle. "Messy" is better. It always is.
> Enable client to setAttribute that is sent once to each region server
> ---------------------------------------------------------------------
>
> Key: HBASE-9291
> URL: https://issues.apache.org/jira/browse/HBASE-9291
> Project: HBase
> Issue Type: New Feature
> Components: IPC/RPC
> Reporter: James Taylor
>
> Currently a Scan and Mutation allow the client to set its own attributes that
> get passed through the RPC layer and are accessible from a coprocessor. This
> is very handy, but breaks down if the amount of information is large, since
> this information ends up being sent again and again to every region. Clients
> can work around this with an endpoint "pre" and "post" coprocessor invocation
> that:
> 1) sends the information and caches it on the region server in the "pre"
> invocation
> 2) invokes the Scan or sends the batch of Mutations, and then
> 3) removes it in the "post" invocation.
> In this case, the client is forced to identify all region servers (ideally,
> all region servers that will be involved in the Scan/Mutation), make extra
> RPC calls, manage the caching of the information on the region server,
> age-out the information (in case the client dies before step (3) that clears
> the cached information), and must deal with the possibility of a split
> occurring while this operation is in-progress.
> Instead, it'd be much better if an attribute could be identified as a "region
> server" attribute in OperationWithAttributes and the HBase RPC layer would
> take care of doing the above.
> The use case where the above are necessary in Phoenix include:
> 1) Hash joins, where the results of the smaller side of a join scan are
> packaged up and sent to each region server, and
> 2) Secondary indexing, where the metadata of knowing a) which column
> family/column qualifier pairs and b) which part of the row key contributes to
> which indexes are sent to each region server that will process a batched put.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
