What are the current limitations on storing large values in Accumulo? Specifically, what would break first and at how big would a value have to be to cause the break? Also, is storing lots of large values much worse than storing occasional large values?

I can of course test this but it'd be nice to hear what is already known about the subject first.

Aaron

It depends on how large loggers' JVM can get.  We typically run them with
less memory than the tablet servers, so we see the loggers break when the
large mutation hits the logger.

If your logger has a large JVM, then your tablet server will crash
receiving the message, or serializing it to the loggers, and then when it
tries to recover after crashing, which will slowly take down your whole
cluster.

However, if a large value is bulk loaded, you will see the tablet servers
crash when they do a major compaction.

The upper limit is something like 1/3 of the JVM memory size, but in
practice, if you have a lot of large values, this limit will be much lower.

-Eric

On Wed, Mar 14, 2012 at 1:00 PM, Aaron Cordova <[EMAIL PROTECTED]> wrote:

> What are the current limitations on storing large values in Accumulo?
> Specifically, what would break first and at how big would a value have to
> be to cause the break? Also, is storing lots of large values much worse
> than storing occasional large values?
>
> I can of course test this but it'd be nice to hear what is already known
> about the subject first.
>
> Aaron

Concurrency is another consideration.  There could be multiple large
values in flight at any given time.  For example multiple clients
could be writing mutations with large values, while minor compactions
are writing out large values, and major compactions reading and
writing large values.  In this case all of these threads running
concurrently could exhaust memory when no single thread would.  The
probability of this happening at the same time increases as more large
values are written.

Keith

On Wed, Mar 14, 2012 at 1:22 PM, Eric Newton <[EMAIL PROTECTED]> wrote:
> It depends on how large loggers' JVM can get.  We typically run them with
> less memory than the tablet servers, so we see the loggers break when the
> large mutation hits the logger.
>
> If your logger has a large JVM, then your tablet server will crash receiving
> the message, or serializing it to the loggers, and then when it tries to
> recover after crashing, which will slowly take down your whole cluster.
>
> However, if a large value is bulk loaded, you will see the tablet servers
> crash when they do a major compaction.
>
> The upper limit is something like 1/3 of the JVM memory size, but in
> practice, if you have a lot of large values, this limit will be much lower.
>
> -Eric
>
>
> On Wed, Mar 14, 2012 at 1:00 PM, Aaron Cordova <[EMAIL PROTECTED]> wrote:
>>
>> What are the current limitations on storing large values in Accumulo?
>> Specifically, what would break first and at how big would a value have to be
>> to cause the break? Also, is storing lots of large values much worse than
>> storing occasional large values?
>>
>> I can of course test this but it'd be nice to hear what is already known
>> about the subject first.
>>
>> Aaron
>
>