You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@hbase.apache.org by Eric Czech <er...@nextbigsound.com> on 2012/08/25 23:44:24 UTC
MemStore and prefix encoding
Hi everyone,
Does prefix encoding apply to rows in MemStores or does it only apply
to rows on disk in HFiles? I'm trying to decide if I should still
favor larger values in order to not repeat keys, column families, and
qualifiers more than necessary and while prefix encoding seems to
negate that concern for storage on disk, I'm not sure if it's still
applicable to in-memory storage.
Also, I had two other quick (unrelated) questions and I assume it'd be
less annoying if I put them all in one email:
1. Do column families defined for a table introduce any overhead for
rows that don't put any values in them? I don't think that's the case
but I wanted to be sure.
2. Do writes with the same key, column family, qualifier, and
timestamp count towards the version limit?
Thanks for the help!
Re: MemStore and prefix encoding
Posted by Stack <st...@duboce.net>.
On Tue, Aug 28, 2012 at 9:59 AM, Joe Pallas <pa...@cs.stanford.edu> wrote:
>
> On Aug 25, 2012, at 2:57 PM, lars hofhansl wrote:
>
>> Each column family is its own store. All stores are flushed together, so have many add overhead (especially if a few tend to hold a lot of data, but the others don't, leading to very many small store files that need to be compacted).
>
> If there are no writes at all to a CF, will a flush still create a new store file? I’m thinking about a case where one CF is basically write-once and another CF gets frequent updates.
>
I was going to say we won't write a file if no kvs, but looking at
code, it looks like we could write a file w/ nothing but metadata:
http://hbase.apache.org/xref/org/apache/hadoop/hbase/regionserver/Store.html#725
I see a check up in HRegion where we look at memstore size, the
aggregated size of all CFs, and if it is null we'll skip a flush but I
did not see an equivalent check at the CF level (after three minutes
of looking). We should add one if not one already.
St.Ack
Re: MemStore and prefix encoding
Posted by Joe Pallas <pa...@cs.stanford.edu>.
On Aug 25, 2012, at 2:57 PM, lars hofhansl wrote:
> Each column family is its own store. All stores are flushed together, so have many add overhead (especially if a few tend to hold a lot of data, but the others don't, leading to very many small store files that need to be compacted).
If there are no writes at all to a CF, will a flush still create a new store file? I’m thinking about a case where one CF is basically write-once and another CF gets frequent updates.
Thanks.
joe
Re: MemStore and prefix encoding
Posted by Enis Söztutar <en...@hortonworks.com>.
I would still caution relying on the sorting order between values of the
same cf, qualifier and timestamp. If for example, there is a Delete, it
will eclipse subsequent Puts given the same timestamp, even though Put
happened after Delete.
Enis
On Mon, Aug 27, 2012 at 9:20 AM, Tom Brown <to...@gmail.com> wrote:
> Lars,
>
> I have been relying on the expected behavior (if I write another cell
> with the same {key, family, qualifier, version} it won't return the
> previous one) so you're answer was confusing to me. I did more
> research and I found that the HBase guide specifies that behavior (see
> section 5.8.1 of http://hbase.apache.org/book.html).
>
> Have I misunderstood something? Can I rely on behavior that is
> specified in the guide?
>
> Thanks again!
>
> --Tom
>
> On Sun, Aug 26, 2012 at 6:43 AM, Eric Czech <er...@nextbigsound.com> wrote:
> > Thanks for the info lars!
> >
> > In the potential use case I have for writing at the same timestamp,
> > the values would always be the same anyways so I should be good.
> >
> > On Sat, Aug 25, 2012 at 9:12 PM, lars hofhansl <lh...@yahoo.com>
> wrote:
> >> I checked the code to be sure...
> >>
> >>
> >> In ScanWildcardColumnTracker we have this:
> >>
> >> if (sameAsPreviousTSAndType(timestamp, type)) {
> >> return ScanQueryMatcher.MatchCode.SKIP;
> >> }
> >>
> >>
> >> And in ExplicitColumnTracker there is this:
> >>
> >> if (sameAsPreviousTS(timestamp)) {
> >> //If duplicate, skip this Key
> >> return ScanQueryMatcher.MatchCode.SKIP;
> >> }
> >>
> >>
> >> I.e. the first KV is kept and the subsequent ones (with the same TS)
> are skipped.
> >>
> >> My point remains, though: Do not rely on this.
> >> (Though it will probably stay the way it is, because that is the most
> efficient way to handle this in forward only scanners.)
> >>
> >> -- Lars
> >>
> >>
> >>
> >> ________________________________
> >> From: Tom Brown <to...@gmail.com>
> >> To: "user@hbase.apache.org" <us...@hbase.apache.org>; lars hofhansl <
> lhofhansl@yahoo.com>
> >> Sent: Saturday, August 25, 2012 4:54 PM
> >> Subject: Re: MemStore and prefix encoding
> >>
> >>
> >> I thought when multiple values with the same key, family, qualifier and
> timestamps were written, the one that was written latest (as determined by
> position in the store) would be read. Is that not the case?
> >>
> >> --Tom
> >>
> >> On Saturday, August 25, 2012, lars hofhansl <lh...@yahoo.com>
> wrote:
> >>> The prefix encoding applies to blocks in the HFiles and in the block
> cache, but not to the memstore.
> >>>
> >>>
> >>> #1 Yes. Each column family is its own store. All stores are flushed
> together, so have many add overhead (especially if a few tend to hold a lot
> of data, but the others don't, leading to very many small store files that
> need to be compacted).
> >>> #2 There is only one key with the same key, column family, qualifier,
> and timestamp (if you write multiple with the same timestamp it is
> undefined which one you'll get back when you read the next time). So that
> does not make sense. Writes with the same key, column family, qualifier
> (each with a different timestamp) count towards the version limit.
> >>>
> >>> -- Lars
> >>>
> >>>
> >>> ----- Original Message -----
> >>> From: Eric Czech <er...@nextbigsound.com>
> >>> To: user <us...@hbase.apache.org>
> >>> Cc:
> >>> Sent: Saturday, August 25, 2012 2:44 PM
> >>> Subject: MemStore and prefix encoding
> >>>
> >>> Hi everyone,
> >>>
> >>> Does prefix encoding apply to rows in MemStores or does it only apply
> >>> to rows on disk in HFiles? I'm trying to decide if I should still
> >>> favor larger values in order to not repeat keys, column families, and
> >>> qualifiers more than necessary and while prefix encoding seems to
> >>> negate that concern for storage on disk, I'm not sure if it's still
> >>> applicable to in-memory storage.
> >>>
> >>> Also, I had two other quick (unrelated) questions and I assume it'd be
> >>> less annoying if I put them all in one email:
> >>>
> >>> 1. Do column families defined for a table introduce any overhead for
> >>> rows that don't put any values in them? I don't think that's the case
> >>> but I wanted to be sure.
> >>>
> >>> 2. Do writes with the same key, column family, qualifier, and
> >>> timestamp count towards the version limit?
> >>>
> >>> Thanks for the help!
> >>>
> >>>
>
Re: MemStore and prefix encoding
Posted by Stack <st...@duboce.net>.
On Mon, Aug 27, 2012 at 9:20 AM, Tom Brown <to...@gmail.com> wrote:
> Lars,
>
> I have been relying on the expected behavior (if I write another cell
> with the same {key, family, qualifier, version} it won't return the
> previous one) so you're answer was confusing to me. I did more
> research and I found that the HBase guide specifies that behavior (see
> section 5.8.1 of http://hbase.apache.org/book.html).
>
> Have I misunderstood something? Can I rely on behavior that is
> specified in the guide?
>
If the code AND the refguide say the same thing, apart from that being
a minor miracle, I'd say its unlikely the behavior will change, not
w/o really good reason (We don't etch anything in stone around these
parts).
St.Ack
Re: MemStore and prefix encoding
Posted by Tom Brown <to...@gmail.com>.
Lars,
I have been relying on the expected behavior (if I write another cell
with the same {key, family, qualifier, version} it won't return the
previous one) so you're answer was confusing to me. I did more
research and I found that the HBase guide specifies that behavior (see
section 5.8.1 of http://hbase.apache.org/book.html).
Have I misunderstood something? Can I rely on behavior that is
specified in the guide?
Thanks again!
--Tom
On Sun, Aug 26, 2012 at 6:43 AM, Eric Czech <er...@nextbigsound.com> wrote:
> Thanks for the info lars!
>
> In the potential use case I have for writing at the same timestamp,
> the values would always be the same anyways so I should be good.
>
> On Sat, Aug 25, 2012 at 9:12 PM, lars hofhansl <lh...@yahoo.com> wrote:
>> I checked the code to be sure...
>>
>>
>> In ScanWildcardColumnTracker we have this:
>>
>> if (sameAsPreviousTSAndType(timestamp, type)) {
>> return ScanQueryMatcher.MatchCode.SKIP;
>> }
>>
>>
>> And in ExplicitColumnTracker there is this:
>>
>> if (sameAsPreviousTS(timestamp)) {
>> //If duplicate, skip this Key
>> return ScanQueryMatcher.MatchCode.SKIP;
>> }
>>
>>
>> I.e. the first KV is kept and the subsequent ones (with the same TS) are skipped.
>>
>> My point remains, though: Do not rely on this.
>> (Though it will probably stay the way it is, because that is the most efficient way to handle this in forward only scanners.)
>>
>> -- Lars
>>
>>
>>
>> ________________________________
>> From: Tom Brown <to...@gmail.com>
>> To: "user@hbase.apache.org" <us...@hbase.apache.org>; lars hofhansl <lh...@yahoo.com>
>> Sent: Saturday, August 25, 2012 4:54 PM
>> Subject: Re: MemStore and prefix encoding
>>
>>
>> I thought when multiple values with the same key, family, qualifier and timestamps were written, the one that was written latest (as determined by position in the store) would be read. Is that not the case?
>>
>> --Tom
>>
>> On Saturday, August 25, 2012, lars hofhansl <lh...@yahoo.com> wrote:
>>> The prefix encoding applies to blocks in the HFiles and in the block cache, but not to the memstore.
>>>
>>>
>>> #1 Yes. Each column family is its own store. All stores are flushed together, so have many add overhead (especially if a few tend to hold a lot of data, but the others don't, leading to very many small store files that need to be compacted).
>>> #2 There is only one key with the same key, column family, qualifier, and timestamp (if you write multiple with the same timestamp it is undefined which one you'll get back when you read the next time). So that does not make sense. Writes with the same key, column family, qualifier (each with a different timestamp) count towards the version limit.
>>>
>>> -- Lars
>>>
>>>
>>> ----- Original Message -----
>>> From: Eric Czech <er...@nextbigsound.com>
>>> To: user <us...@hbase.apache.org>
>>> Cc:
>>> Sent: Saturday, August 25, 2012 2:44 PM
>>> Subject: MemStore and prefix encoding
>>>
>>> Hi everyone,
>>>
>>> Does prefix encoding apply to rows in MemStores or does it only apply
>>> to rows on disk in HFiles? I'm trying to decide if I should still
>>> favor larger values in order to not repeat keys, column families, and
>>> qualifiers more than necessary and while prefix encoding seems to
>>> negate that concern for storage on disk, I'm not sure if it's still
>>> applicable to in-memory storage.
>>>
>>> Also, I had two other quick (unrelated) questions and I assume it'd be
>>> less annoying if I put them all in one email:
>>>
>>> 1. Do column families defined for a table introduce any overhead for
>>> rows that don't put any values in them? I don't think that's the case
>>> but I wanted to be sure.
>>>
>>> 2. Do writes with the same key, column family, qualifier, and
>>> timestamp count towards the version limit?
>>>
>>> Thanks for the help!
>>>
>>>
Re: MemStore and prefix encoding
Posted by Eric Czech <er...@nextbigsound.com>.
Thanks for the info lars!
In the potential use case I have for writing at the same timestamp,
the values would always be the same anyways so I should be good.
On Sat, Aug 25, 2012 at 9:12 PM, lars hofhansl <lh...@yahoo.com> wrote:
> I checked the code to be sure...
>
>
> In ScanWildcardColumnTracker we have this:
>
> if (sameAsPreviousTSAndType(timestamp, type)) {
> return ScanQueryMatcher.MatchCode.SKIP;
> }
>
>
> And in ExplicitColumnTracker there is this:
>
> if (sameAsPreviousTS(timestamp)) {
> //If duplicate, skip this Key
> return ScanQueryMatcher.MatchCode.SKIP;
> }
>
>
> I.e. the first KV is kept and the subsequent ones (with the same TS) are skipped.
>
> My point remains, though: Do not rely on this.
> (Though it will probably stay the way it is, because that is the most efficient way to handle this in forward only scanners.)
>
> -- Lars
>
>
>
> ________________________________
> From: Tom Brown <to...@gmail.com>
> To: "user@hbase.apache.org" <us...@hbase.apache.org>; lars hofhansl <lh...@yahoo.com>
> Sent: Saturday, August 25, 2012 4:54 PM
> Subject: Re: MemStore and prefix encoding
>
>
> I thought when multiple values with the same key, family, qualifier and timestamps were written, the one that was written latest (as determined by position in the store) would be read. Is that not the case?
>
> --Tom
>
> On Saturday, August 25, 2012, lars hofhansl <lh...@yahoo.com> wrote:
>> The prefix encoding applies to blocks in the HFiles and in the block cache, but not to the memstore.
>>
>>
>> #1 Yes. Each column family is its own store. All stores are flushed together, so have many add overhead (especially if a few tend to hold a lot of data, but the others don't, leading to very many small store files that need to be compacted).
>> #2 There is only one key with the same key, column family, qualifier, and timestamp (if you write multiple with the same timestamp it is undefined which one you'll get back when you read the next time). So that does not make sense. Writes with the same key, column family, qualifier (each with a different timestamp) count towards the version limit.
>>
>> -- Lars
>>
>>
>> ----- Original Message -----
>> From: Eric Czech <er...@nextbigsound.com>
>> To: user <us...@hbase.apache.org>
>> Cc:
>> Sent: Saturday, August 25, 2012 2:44 PM
>> Subject: MemStore and prefix encoding
>>
>> Hi everyone,
>>
>> Does prefix encoding apply to rows in MemStores or does it only apply
>> to rows on disk in HFiles? I'm trying to decide if I should still
>> favor larger values in order to not repeat keys, column families, and
>> qualifiers more than necessary and while prefix encoding seems to
>> negate that concern for storage on disk, I'm not sure if it's still
>> applicable to in-memory storage.
>>
>> Also, I had two other quick (unrelated) questions and I assume it'd be
>> less annoying if I put them all in one email:
>>
>> 1. Do column families defined for a table introduce any overhead for
>> rows that don't put any values in them? I don't think that's the case
>> but I wanted to be sure.
>>
>> 2. Do writes with the same key, column family, qualifier, and
>> timestamp count towards the version limit?
>>
>> Thanks for the help!
>>
>>
Re: MemStore and prefix encoding
Posted by lars hofhansl <lh...@yahoo.com>.
I checked the code to be sure...
In ScanWildcardColumnTracker we have this:
if (sameAsPreviousTSAndType(timestamp, type)) {
return ScanQueryMatcher.MatchCode.SKIP;
}
And in ExplicitColumnTracker there is this:
if (sameAsPreviousTS(timestamp)) {
//If duplicate, skip this Key
return ScanQueryMatcher.MatchCode.SKIP;
}
I.e. the first KV is kept and the subsequent ones (with the same TS) are skipped.
My point remains, though: Do not rely on this.
(Though it will probably stay the way it is, because that is the most efficient way to handle this in forward only scanners.)
-- Lars
________________________________
From: Tom Brown <to...@gmail.com>
To: "user@hbase.apache.org" <us...@hbase.apache.org>; lars hofhansl <lh...@yahoo.com>
Sent: Saturday, August 25, 2012 4:54 PM
Subject: Re: MemStore and prefix encoding
I thought when multiple values with the same key, family, qualifier and timestamps were written, the one that was written latest (as determined by position in the store) would be read. Is that not the case?
--Tom
On Saturday, August 25, 2012, lars hofhansl <lh...@yahoo.com> wrote:
> The prefix encoding applies to blocks in the HFiles and in the block cache, but not to the memstore.
>
>
> #1 Yes. Each column family is its own store. All stores are flushed together, so have many add overhead (especially if a few tend to hold a lot of data, but the others don't, leading to very many small store files that need to be compacted).
> #2 There is only one key with the same key, column family, qualifier, and timestamp (if you write multiple with the same timestamp it is undefined which one you'll get back when you read the next time). So that does not make sense. Writes with the same key, column family, qualifier (each with a different timestamp) count towards the version limit.
>
> -- Lars
>
>
> ----- Original Message -----
> From: Eric Czech <er...@nextbigsound.com>
> To: user <us...@hbase.apache.org>
> Cc:
> Sent: Saturday, August 25, 2012 2:44 PM
> Subject: MemStore and prefix encoding
>
> Hi everyone,
>
> Does prefix encoding apply to rows in MemStores or does it only apply
> to rows on disk in HFiles? I'm trying to decide if I should still
> favor larger values in order to not repeat keys, column families, and
> qualifiers more than necessary and while prefix encoding seems to
> negate that concern for storage on disk, I'm not sure if it's still
> applicable to in-memory storage.
>
> Also, I had two other quick (unrelated) questions and I assume it'd be
> less annoying if I put them all in one email:
>
> 1. Do column families defined for a table introduce any overhead for
> rows that don't put any values in them? I don't think that's the case
> but I wanted to be sure.
>
> 2. Do writes with the same key, column family, qualifier, and
> timestamp count towards the version limit?
>
> Thanks for the help!
>
>
Re: MemStore and prefix encoding
Posted by Tom Brown <to...@gmail.com>.
I thought when multiple values with the same key, family, qualifier and
timestamps were written, the one that was written latest (as determined by
position in the store) would be read. Is that not the case?
--Tom
On Saturday, August 25, 2012, lars hofhansl <lh...@yahoo.com> wrote:
> The prefix encoding applies to blocks in the HFiles and in the block
cache, but not to the memstore.
>
>
> #1 Yes. Each column family is its own store. All stores are flushed
together, so have many add overhead (especially if a few tend to hold a lot
of data, but the others don't, leading to very many small store files that
need to be compacted).
> #2 There is only one key with the same key, column family, qualifier, and
timestamp (if you write multiple with the same timestamp it is undefined
which one you'll get back when you read the next time). So that does not
make sense. Writes with the same key, column family, qualifier (each with a
different timestamp) count towards the version limit.
>
> -- Lars
>
>
> ----- Original Message -----
> From: Eric Czech <er...@nextbigsound.com>
> To: user <us...@hbase.apache.org>
> Cc:
> Sent: Saturday, August 25, 2012 2:44 PM
> Subject: MemStore and prefix encoding
>
> Hi everyone,
>
> Does prefix encoding apply to rows in MemStores or does it only apply
> to rows on disk in HFiles? I'm trying to decide if I should still
> favor larger values in order to not repeat keys, column families, and
> qualifiers more than necessary and while prefix encoding seems to
> negate that concern for storage on disk, I'm not sure if it's still
> applicable to in-memory storage.
>
> Also, I had two other quick (unrelated) questions and I assume it'd be
> less annoying if I put them all in one email:
>
> 1. Do column families defined for a table introduce any overhead for
> rows that don't put any values in them? I don't think that's the case
> but I wanted to be sure.
>
> 2. Do writes with the same key, column family, qualifier, and
> timestamp count towards the version limit?
>
> Thanks for the help!
>
>
Re: MemStore and prefix encoding
Posted by lars hofhansl <lh...@yahoo.com>.
The prefix encoding applies to blocks in the HFiles and in the block cache, but not to the memstore.
#1 Yes. Each column family is its own store. All stores are flushed together, so have many add overhead (especially if a few tend to hold a lot of data, but the others don't, leading to very many small store files that need to be compacted).
#2 There is only one key with the same key, column family, qualifier, and timestamp (if you write multiple with the same timestamp it is undefined which one you'll get back when you read the next time). So that does not make sense. Writes with the same key, column family, qualifier (each with a different timestamp) count towards the version limit.
-- Lars
----- Original Message -----
From: Eric Czech <er...@nextbigsound.com>
To: user <us...@hbase.apache.org>
Cc:
Sent: Saturday, August 25, 2012 2:44 PM
Subject: MemStore and prefix encoding
Hi everyone,
Does prefix encoding apply to rows in MemStores or does it only apply
to rows on disk in HFiles? I'm trying to decide if I should still
favor larger values in order to not repeat keys, column families, and
qualifiers more than necessary and while prefix encoding seems to
negate that concern for storage on disk, I'm not sure if it's still
applicable to in-memory storage.
Also, I had two other quick (unrelated) questions and I assume it'd be
less annoying if I put them all in one email:
1. Do column families defined for a table introduce any overhead for
rows that don't put any values in them? I don't think that's the case
but I wanted to be sure.
2. Do writes with the same key, column family, qualifier, and
timestamp count towards the version limit?
Thanks for the help!