And it will probably evict everyone else that was already present.
Hello latency.

J-D

On Thu, Nov 17, 2011 at 2:08 PM, lars hofhansl <[email protected]> wrote:
> Hi Sam,
> The idea is that the entire result of the scan will not fit into the cache if 
> the scan scans a "reasonable" number of cells, and hence it unlikely that 
> another scan will hit cached blocks before they get evicted, especially when 
> using an LRU cache.
>
> -- Lars
>
>
> ----- Original Message -----
> From: Sam Seigal <[email protected]>
> To: [email protected]
> Cc:
> Sent: Thursday, November 17, 2011 1:44 PM
> Subject: block caching
>
> I have a table that I only use for generating indexes. It rarely will
> have random reads, but will have M/R jobs running against it
> constantly for generating indexes. Even the index table, random reads
> will be rare. It will mostly be used for scanning blocks of data.
>
>
> According to HBase The Definitive Guide
>
> "As HBase reads entire blocks of data for efficient IO usage it
> retains these blocks in an in-memory cache, so that subsequent reads
> do not need any disk operation. The default of true enables the block
> cache for every read operation. But if your use-case only ever has
> sequential reads on a particular column family it is advisable to
> disable it from polluting the block cache by setting the block cache
> enabled flag to false. "
>
> "There are other options you can use to influence how the block cache
> is used, for example during a scan operation. This is useful during
> full table scans so that you do not cause a major churn on the cache.
> See the section called “Configuration” for more information about this
> feature."
>
> "Scan instances can be set to use the block cache in the region server
> via the setCacheBlocks() method. For scans used with MapReduce jobs,
> this should be false. For frequently accessed rows, it is advisable to
> use the block cache."
>
>
> What is the reasoning behind the above ?  Why is using a block cache
> for M/R jobs not a good idea if it is doing full table scans ?
>
>

Reply via email to