incubator-cassandra-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Boris Yen <yulin...@gmail.com>
Subject Re: Possibility of going OOM using get_count
Date Mon, 26 Sep 2011 02:32:56 GMT
Hi Aaron,

Thanks for the explanation, I know the performance will be varied when the
offset is a very large number, just like what has been mentioned
on CASSANDRA-261. Even if the users implement the offset on the client side,
they suffer the same issues, I just think it would be nice if cassandra can
provide this function internally, of course this function will have its
limitation, just like any other functions cassnadra has, Counter, for
example.

In CASSANDRA-261, it seems cassandra had the offset function, however, due
to the some RR issues it has been removed, CASSANDRA-286. I think the reason
why CASSANDRA-261 has the RR issue is because it changes the internal
mechanism in order to provide the offset function. Unlike
CASSANDRA-2894<https://issues.apache.org/jira/browse/CASSANDRA-2894>,
it only changes code in the "CassandraServer", it should not have the same
issue as CASSANDRA-261. Therefore, I was wondering if you could re-consider
to put the offset function back to cassandra. It should be really helpful
for many users.

Regards
Boris

On Sun, Sep 25, 2011 at 12:21 PM, aaron morton <aaron@thelastpickle.com>wrote:

> The changes in get_count() are designed to stop counts for very large rows
> running out of memory as they try to hold millions of columns in memory.
>
> So if you ask to count all the cols in a row with 1M cols, it will (by
> default) read the first 1024 columns, and then the next 1024 using the last
> column read as the first column for the next page.
>
> The important part is that it is actually reading the columns. Tombstones
> mean we do not know if a column should be a member of the result set for a
> query until it is read and reconciled with all the other versions of a
> column. e.g. 3 sstables have each have a value for a column, if one is a
> tombstone then the column may or may not be deleted. We do not know until
> all 3 column versions are reconciled.
>
> get_count() is like get_slice() but we do not return the columns, just the
> count of them. Counting 1M columns still takes a long time. And find the
> 999,980th column will also take a long time, but if you know the name of the
> 999,980th column it will be mucho faster.
>
> Some experiments I did a while ago on query plans
> http://thelastpickle.com/2011/07/04/Cassandra-Query-Plans/ - cass 1.0 will
> probably invalidate this.
>
> Cheers
>
>
> -----------------
> Aaron Morton
> Freelance Cassandra Developer
> @aaronmorton
> http://www.thelastpickle.com
>
> On 23/09/2011, at 6:01 PM, Boris Yen wrote:
>
>
>
> On Fri, Sep 23, 2011 at 12:28 PM, aaron morton <aaron@thelastpickle.com>wrote:
>
>> Offsets have been discussed in previously. IIRC the main concerns were
>> either:
>>
>> There is no way to reliably count to start the offset, i.e. we do not lock
>> the row
>>
>
> In the new get_count function, cassandra does the internal paging in order
> to get the total count. Without locking the row,  the count could still be
> unreliable (someone might be deleting some columns while cassandra is
> counting the columns).
>
>
>>
>> Or performance related in, as there is not a reliable way to skip 10,000
>> columns other than counting 10,000 columns. With a start col we can search.
>>
>>
> I am just curious, basically "skip 10,000 columns to get the start column"
> can be done as what cassandra does for new get_count function (internal
> paging). I just can not think of a reason why it is doable for get_count but
> it can not be done for the offset.
>
> I know the result might not be reliable and the performance might be varied
> depends on the offset, but if cassandra can using internal paging to get
> count, it should be able the apply the same method to get the start column
> for the offset.
>
>
>> Cheers
>>
>>  -----------------
>> Aaron Morton
>> Freelance Cassandra Developer
>> @aaronmorton
>> http://www.thelastpickle.com
>>
>> On 22/09/2011, at 8:50 PM, Boris Yen wrote:
>>
>> I was wondering if it is possible to use similar way as CASSANDRA-2894<https://issues.apache.org/jira/browse/CASSANDRA-2894>
to
>> have the slice_predict support the offset concept? With the offset, it would
>> be much easier to implement the paging from the client side.
>>
>> Boris
>>
>> On Mon, Sep 19, 2011 at 9:45 PM, Jonathan Ellis <jbellis@gmail.com>wrote:
>>
>>> Unfortunately no, because you don't know what the actual
>>> last-column-counted was.
>>>
>>> On Mon, Sep 19, 2011 at 4:25 AM, aaron morton <aaron@thelastpickle.com>
>>> wrote:
>>> > get_count() supports the same predicate as get_slice. So you can
>>> implement
>>> > the paging yourself.
>>> > Cheers
>>> > -----------------
>>> > Aaron Morton
>>> > Freelance Cassandra Developer
>>> > @aaronmorton
>>> > http://www.thelastpickle.com
>>> > On 19/09/2011, at 8:45 PM, Tharindu Mathew wrote:
>>> >
>>> >
>>> > On Mon, Sep 19, 2011 at 12:40 PM, Benoit Perroud <benoit@noisette.ch>
>>> wrote:
>>> >>
>>> >> The workaround for 0.7 is calling get_slice and count on client side.
>>> >> It's heavier, sure, but you will then be able to set start column
>>> >> accordingly.
>>> >
>>> > I was afraid of that :(
>>> > Will follow that method. Thanks.
>>> >>
>>> >>
>>> >> 2011/9/19 Tharindu Mathew <mccloud35@gmail.com>:
>>> >> > Thanks Aaron and Jake for the replies.
>>> >> > Any chance of a possible workaround to use for Cassandra 0.7?
>>> >> >
>>> >> > On Mon, Sep 19, 2011 at 3:48 AM, aaron morton <
>>> aaron@thelastpickle.com>
>>> >> > wrote:
>>> >> >>
>>> >> >> Cool
>>> >> >> Thanks, A
>>> >> >> -----------------
>>> >> >> Aaron Morton
>>> >> >> Freelance Cassandra Developer
>>> >> >> @aaronmorton
>>> >> >> http://www.thelastpickle.com
>>> >> >> On 19/09/2011, at 9:55 AM, Jake Luciani wrote:
>>> >> >>
>>> >> >> This is fixed in 1.0
>>> >> >> https://issues.apache.org/jira/browse/CASSANDRA-2894
>>> >> >>
>>> >> >> On Sun, Sep 18, 2011 at 2:16 PM, Tharindu Mathew <
>>> mccloud35@gmail.com>
>>> >> >> wrote:
>>> >> >>>
>>> >> >>> Hi everyone,
>>> >> >>> I noticed this line in the API docs,
>>> >> >>>
>>> >> >>> The method is not O(1). It takes all the columns from disk
to
>>> >> >>> calculate
>>> >> >>> the answer. The only benefit of the method is that you
do not need
>>> to
>>> >> >>> pull
>>> >> >>> all the columns over Thrift interface to count them.
>>> >> >>>
>>> >> >>> Does this mean if a row has a large number of columns calling
this
>>> >> >>> method
>>> >> >>> might make it go OOM?
>>> >> >>> Thanks in advance.
>>> >> >>> --
>>> >> >>> Regards,
>>> >> >>>
>>> >> >>> Tharindu
>>> >> >>> blog: http://mackiemathew.com/
>>> >> >>
>>> >> >>
>>> >> >>
>>> >> >> --
>>> >> >> http://twitter.com/tjake
>>> >> >>
>>> >> >
>>> >> >
>>> >> >
>>> >> > --
>>> >> > Regards,
>>> >> >
>>> >> > Tharindu
>>> >> > blog: http://mackiemathew.com/
>>> >> >
>>> >
>>> >
>>> >
>>> > --
>>> > Regards,
>>> >
>>> > Tharindu
>>> > blog: http://mackiemathew.com/
>>> >
>>> >
>>>
>>>
>>>
>>> --
>>> Jonathan Ellis
>>> Project Chair, Apache Cassandra
>>> co-founder of DataStax, the source for professional Cassandra support
>>> http://www.datastax.com
>>>
>>
>>
>>
>
>

Mime
View raw message