hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From tobe <tobeg3oo...@gmail.com>
Subject Re: Should scan check the limitation of the number of versions?
Date Tue, 26 Aug 2014 11:56:38 GMT
Thanks @nicolas, @andrew and @lars. The problem like this always comes down
to "by design". It depends on the semantic that HBase provides.

As a user, I don't expect different results when I sent the same request at
the same time. I don't care about how HBase operates and I think the
process is determinate and predicable. So if you have to say it depends on
whether the server runs compactions or not, I prefer to a more determinate
semantic.


On Tue, Aug 26, 2014 at 4:54 PM, Nicolas Liochon <nkeywal@gmail.com> wrote:

> (moving to user)
>
> In your first scenario (put "table", "row1", "cf:a", "value1", 100 then put
> "table", "row1", "cf:a", "value1", 200), there is no deletion, so the
> setting KEEP_DELETED_CELLS is not used at all
> The behavior you describe is "as expected": there are two versions until
> the compaction occurs and removes the version not needed, depending on the
> configuration.
> There are some optimizations around this: we skip reading early if the
> timestamps of what we're reading is not in the scan range. So we don't know
> if there is a newer value.
>
> What's the use case you're looking at?
>
> Nicolas
>
>
>
>
>
>
>
>
>
>
>
>
> On Tue, Aug 26, 2014 at 3:36 AM, tobe <tobeg3oogle@gmail.com> wrote:
>
> > @andrew Actually I don't want to see row in TIMERANGE => [0, 150] because
> > it's the overdue version. Should I set {KEEP_DELETED_CELLS => 'true'}? My
> > problem is that even though I don't keep deleted cells, I will get the
> > result which is not what I expect.
> >
> >
> > On Tue, Aug 26, 2014 at 9:24 AM, Andrew Purtell <apurtell@apache.org>
> > wrote:
> >
> > > On Mon, Aug 25, 2014 at 6:13 PM, tobe <tobeg3oogle@gmail.com> wrote:
> > >
> > > > @lars I have set {KEEP_DELETED_CELLS => 'false'} in that table. I
> will
> > > get
> > > > the same result before manually running `flush`. You can try the
> > > commands I
> > > > gave and it's 100% repro.
> > > >
> > >
> > > ​You need KEEP_DELETED_CELLS => 'true'. ​
> > >
> > >
> > >
> > > On Mon, Aug 25, 2014 at 6:13 PM, tobe <tobeg3oogle@gmail.com> wrote:
> > >
> > > > @lars I have set {KEEP_DELETED_CELLS => 'false'} in that table. I
> will
> > > get
> > > > the same result before manually running `flush`. You can try the
> > > commands I
> > > > gave and it's 100% repro.
> > > >
> > > >
> > > > On Tue, Aug 26, 2014 at 2:20 AM, lars hofhansl <larsh@apache.org>
> > wrote:
> > > >
> > > > > Queries of past time ranges only work correctly when
> > KEEP_DELETED_CELLS
> > > > is
> > > > > enabled for the column families.
> > > > >
> > > > >
> > > > > ________________________________
> > > > >  From: tobe <tobeg3oogle@gmail.com>
> > > > > To: hbase-dev <dev@hbase.apache.org>
> > > > > Cc: "user@hbase.apache.org" <user@hbase.apache.org>
> > > > > Sent: Monday, August 25, 2014 4:32 AM
> > > > > Subject: Re: Should scan check the limitation of the number of
> > > versions?
> > > > >
> > > > >
> > > > > I haven't read the code deeply but I have an idea(not sure whether
> > it's
> > > > > right or not). When we scan the the columns, we will skip the one
> > which
> > > > > doesn't match(deleted). Can we use a counter to record this? For
> each
> > > > skip,
> > > > > we add one until it reaches the restrictive number of versions. But
> > we
> > > > have
> > > > > to consider mvcc and others, which seems more complex.
> > > > >
> > > > >
> > > > >
> > > > >
> > > > >
> > > > > On Mon, Aug 25, 2014 at 5:54 PM, tobe <tobeg3oogle@gmail.com>
> wrote:
> > > > >
> > > > > > So far, I have found two problems about this.
> > > > > >
> > > > > > Firstly, HBase-11675 <
> > > > https://issues.apache.org/jira/browse/HBASE-11675
> > > > > >.
> > > > > > It's a little tricky and rarely happens. But it asks users to
be
> > > > careful
> > > > > of
> > > > > > compaction which occurs on server side. They may get different
> > > results
> > > > > > before and after the major compaction.
> > > > > >
> > > > > > Secondly, if you put a value with timestamp 100, then put another
> > > value
> > > > > on
> > > > > > the same column with timestamp 200. Here we set the number of
> > version
> > > > as
> > > > > 1.
> > > > > > So when we get the value of this column, we will get the latest
> one
> > > > with
> > > > > > timestamp 200 and that's right. But if I get with a timerange
> form
> > 0
> > > to
> > > > > > 150, I may get the first value with timestamp 100 before
> compaction
> > > > > > happens. And after compaction happens, you will never get this
> > value
> > > > even
> > > > > > you run the same command.
> > > > > >
> > > > > > It's easy to repro, follow this steps:
> > > > > > hbase(main):001:0> create "table", "cf"
> > > > > > hbase(main):003:0> put "table", "row1", "cf:a", "value1",
100
> > > > > > hbase(main):003:0> put "table", "row1", "cf:a", "value1",
200
> > > > > > hbase(main):026:0> get "table", "row1", {TIMERANGE =>
[0, 150]}
> //
> > > > > before
> > > > > > flush
> > > > > >    row1      column=cf:a, timestamp=100, value=value1
> > > > > > hbase(main):060:0> flush "table"
> > > > > > hbase(main):082:0> get "table", "row1", {TIMERANGE =>
[0, 150]}
> //
> > > > after
> > > > > > flush
> > > > > >    0 row(s) in 0.0050 seconds
> > > > > >
> > > > > > I think the reason of that is we have three restriction to remove
> > > data:
> > > > > > delete, ttl and versions. Any time we get or scan the data,
we
> will
> > > > check
> > > > > > the delete mark and ttl to make sure it will not return to users.
> > But
> > > > for
> > > > > > versions, we don't check this limitation. Our output relies
on
> the
> > > > > > compaction to cleanup the overdue data. Is it possible to add
> this
> > > > > > condition within scan(get is implemented as scan)?
> > > > > >
> > > > >
> > > >
> > >
> > >
> > >
> > > --
> > > Best regards,
> > >
> > >    - Andy
> > >
> > > Problems worthy of attack prove their worth by hitting back. - Piet
> Hein
> > > (via Tom White)
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message