db-derby-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Mike Matrigali <mikem_...@sbcglobal.net>
Subject Re: [jira] Created: (DERBY-510) DERBY-132 resolved ? Table not automatically compressed
Date Tue, 16 Aug 2005 22:53:47 GMT
thanks, this is useful.

It turns out that the compressions metric is not too hard, as the
internal space vti already basically does the work - at least at
a page level granularity.  I think there may be some use for an
admin thread to maintain info about this compression metric over time
and then use that info to decide to do the compression.  For instance
if a table total size is staying constant but the free space is
growing and shrinking over time then there is no need to compress -
similarly if the table is growing and free space is staying constant
again no need to compress.  Another issue is picking
 a good default for how much free space is enough: any space, 10%, 20%,
1 meg, 10 meg, ...

As you point out, scheduling a compression should be based on whether
it will compress.  The other problem that is similar is that currently
there is not way to update the cardinality statistic.  This statistic
is currently updated only when an index is created explictly or, or
when an index is updated internally as part of a compress table.
Unfortunately a number of applications tend to create empty tables and
indexes and then load the data, so never get this statistic correct.

Keeping with the zero admin goal it would be better to figure out a way
to automatically update this, rather than require an explicit call from
the user.   Currently the code to generate this statistic requires
a scan of the entire index and a compare of every row to the next one -
producing a single statistic for every leading set of columns in an
index.  It is used basically to determine the average number of keys
per value for a given value in an index.  Note that other histogram
type information used by other db's are gathered straight from the
btree, and thus don't require any type of statistic maintenance.

Any good ideas on how to tell when we should update that statistic,
some options include:
    o when the table has grown by X%
    o time based
    o number of rows have changed by X%
    o some sort of sample scheme compared with stored result
    o default for small number of rows, and once when reaches N rows.

Rick Hillegas wrote:
> Continuing to maunder, let me fine-tune this a bit:
> 1) Imagine that, on an ongoing basis we maintain some CompressionMetric,
> which measures whether a given table needs compression/reoptimization.
> Dead space might be part of this metric or not. Time since last
> compression could be part of the metric. The metric could be as crude or
> fancy as we like.
> 2) At some point, based on its CompressionMetric, a table Qualifies for
> compression/reoptimization.
> 3) At some fairly fine-grained interval, a low priority thread wakes up,
> looks for Qualifying tables, and compresses/reoptimizes them. By
> default, this thread runs in a 0-administration database, but we expose
> a knob for scheduling/disabling the thread.
> Your original proposal is a degenerate case of this approach and maybe
> it's the first solution we implement. However, we can get fancier as we
> need to support bigger datasets.
> Cheers,
> -Rick
> Rick Hillegas wrote:
>> Hi Mike,
>> I like your suggestions that a low priority thread should perform the
>> compressions and that we should expose a knob for disabling this
>> thread. Here are some further suggestions:
>> Compressing all the tables and recalculating all the statistics once a
>> month could cause quite a hiccup for a large database. Maybe we could
>> do something finer grained. For instance, we could try to make it easy
>> to ask some question like "Is more than 20% of this table's space
>> dead?" No doubt there are some tricky issues in maintaining a
>> per-table dead-space counter and in keeping that counter from being a
>> sync point during writes. However, if we could answer a question like
>> that, then we could pay the compression/reoptimization penalty as we
>> go rather than incurring a heavy, monthly lump-sum tax.
>> Cheers,
>> -Rick
>> Mike Matrigali wrote:
>>> Full compression of derby tables is not done automatically, I
>>> am looking for input on how to schedule such an operation.  An
>>> operation like this is going to have a large cpu, i/o, and
>>> possible temporary disk space impact on the rest of the server.
>>> As a zero admin db I think we should figure out some way to
>>> do this automatically, but I think there are a number of
>>> applications which would not be happy with such a performance
>>> impact not under their control.
>>> My initial thoughts are to pick a default time frame, say
>>> once every 30 days to check for table level events like
>>> compression and statistics generation and then execute the operations
>>> at low priority.  Also add some sort of parameter so that
>>> applications could disable the automatic background jobs.
>>> Note that derby does automatically reclaim space from deletes
>>> for subsequent inserts, but the granularity currently is at
>>> a page level.  So deleting every 3rd or 5th row is the worst
>>> case behavior.  The page level decision was a tradeoff as
>>> reclaiming the space is time consuming so did not want to
>>> schedule to work on a row by row basis.  Currently we schedule
>>> the work when all the rows on a page are marked deleted.
>>> Volker Edelmann (JIRA) wrote:
>>>> DERBY-132 resolved ? Table not automatically compressed
>>>> --------------------------------------------------------
>>>>         Key: DERBY-510
>>>>         URL: http://issues.apache.org/jira/browse/DERBY-510
>>>>     Project: Derby
>>>>        Type: Bug
>>>>    Versions:    Environment: JDK 1.4.2, JDK 1.5.0
>>>> Windows XP
>>>>    Reporter: Volker Edelmann
>>>> I tried a test-program that repeatedly inserts a bunch of data into
>>>> 1 table and repeatedly deletes a bunch of data.
>>>>    // table is not empty  when test-program starts
>>>>     derby.executeSelect("select count(*) c from rclvalues");
>>>>   TestQueries.executeBulkInsertAnalyst(derby.getConnection(),
>>>> 2000000); // insert 2.000.000 rows
>>>>        derby.executeDelete("delete from rclvalues where MOD(id, 3) =
>>>> 0");
>>>>   TestQueries.executeBulkInsertAnalyst(derby.getConnection(), 1000000);
>>>>        derby.executeDelete("delete from rclvalues where MOD(id, 5) =
>>>> 0");
>>>>     derby.executeSelect("select count(*) c from rclvalues");
>>>> At the end of the operation, the table contains approximately the
>>>> same number of rows. But the size of the database has grown from
>>>> 581 MB to 1.22 GB. From the description of item DERBY-132, I hoped
>>>> that Derby does the compression now ( version 10.1.X.X.).

View raw message