Nicolas Spiegelberg created HBASE-8423:
--
Summary: Allow Major Compaction to Use Different Compression
Key: HBASE-8423
URL: https://issues.apache.org/jira/browse/HBASE-8423
Project: HBase
Nicolas Spiegelberg created HBASE-8155:
--
Summary: Allow Record Compression for HLogs
Key: HBASE-8155
URL: https://issues.apache.org/jira/browse/HBASE-8155
Project: HBase
Issue Type: New
I will be attending Berlin Buzzwords. I have a presentation entitled
Multi-tenant HBase Solutions at Facebook that I will present there.
On 4/10/12 4:35 AM, Lars George lars.geo...@gmail.com wrote:
Hi,
Any of you going to Buzzwords this year? Just curious.
Lars
We're using a variant of the Online schema update in our 89 production.
There are significant differences because of the master rewrite, so I
can't speak about the stability on trunk. We don't run with online
splitting, so that's also a large variant from the use cases some of you
are trying to
Jonathan, please read HBASE-4896 for help documentation. It's the #2
search result when I searched for 'hbase arc'. It answers both issues you
filed
On 3/22/12 1:45 AM, Jonathan Hsieh j...@cloudera.com wrote:
I wanted to try the arcanist 'arc lint' mechanism but need some help
getting started.
I'm wondering why HDFS security support should be mandatory? Append makes
sense because there's no way to have a durable system without it.
Security is currently an optional feature implemented as an HBase
co-processor (vs core), correct? Is there a problem (other than minor
inconvenience) with
In 0.92, we can handle multiple compactions but we wait on major compact
if minor compactions are ongoing for the same store. It should be
enqueued and work after the minor compaction finishes, it's just not
immediately enqueued. See HBASE-5330 for some discussion on this. There
should be a log
We had the author of LIRS come to Facebook last year to talk about his
algorithm and general benefits. At the time, we were looking at
increasing block cache efficiency. The general consensus was that it
wasn't an exponential perf gain, so we could get bigger wins from
cache-on-write
: vrodio...@carrieriq.com
From: Nicolas Spiegelberg [nspiegelb...@fb.com]
Sent: Tuesday, February 21, 2012 9:01 AM
To: dev@hbase.apache.org
Subject: Re: LIRS cache as an alternative to LRU cache
We had the author of LIRS come to Facebook last year to talk about
:
If it was ARC (which uses both LRU and LFU) we'd have patenting issues
with IBM, what we have is closer to a 2Q:
http://www.vldb.org/conf/1994/P439.PDF
J-D
On Tue, Feb 21, 2012 at 9:19 AM, Nicolas Spiegelberg
nspiegelb...@fb.com wrote:
Vlad,
You're correct. The existing algorithm
1. my CF were already working with BF they used ROWCOL, (i didn't pay
attention to that at the time i wrote my answers)
2. I see form the logs that the BF is already 100% - is it bad? should I
had more memory for BF?
Since Bloom Filters are a probabilistic optimization, it's kinda hard to
analyze
@Stack: this is an 89-master change. It should have that tag. I'll look
to figure out why the precommit hook isn't working properly.
On 2/2/12 8:59 PM, Stack st...@duboce.net wrote:
Mikhail, the below doesn't have the hbase jira in it?
St.Ack
On Thu, Feb 2, 2012 at 3:30 PM,
I think HFile upgrade in particular is more complicated than you think.
We currently have production traffic running with HFileV1. It has a
5-min
SLA. We can't afford to take the entire downtime to rewrite 100GB (or
whatever) worth of data. We need to do this while the cluster is live.
AFAIK
1) The trunk has more master work. Specifically 89-master related
features that we implemented for our grid architecture.
You have a list of issues Nicolas?
We have a list, but it's non-trivial. You're more than welcome to help if
you want. :)
2) Client/Server compatibility. Even if we
Currently, we at Facebook are developing mostly on 89 but also doing some
significant exploratory work on trunk. I think that most of our
development will continue to be done on 89 in the near future. We plan to
release some lower-risk projects on 94. However, we won't even entertain
a
I think it's also good to make a crucial distinction here: it's not like
Facebook has a concrete wide-scale upgrade strategy, these 3 points are
deal-breakers that won't allow us to even entertain an upgrade strategy.
This is just as critical as HDFS data loss was in 0.90: it's something we
can't
Traditionally there have been no guarantees of cross major version
compatibility. RPC especially. Never a rolling upgrade from an 0.90 to an
0.92, for example. For persistent data, there is a migration utility for
upping from one major release to the next.
I'm advocating that RPC compatibility
1) The trunk has more master work. Specifically 89-master related
features that we implemented for our grid architecture.
You have a list of issues Nicolas?
It's at least 3 major features. We're more than willing to work with you
on porting. It sounds like eBay Cloudera have some possible
+1
It's much more visible and bold. I like the Golden Gate reference...
On 1/11/12 8:42 AM, Stack st...@duboce.net wrote:
I attached to https://issues.apache.org/jira/browse/HBASE-5115 what it
looks like. International Orange (Engineering) is the color of the
golden gate [1], and is a strong
Mikael,
Hi, I wrote the current compaction algorithm, so I should be able to
answer most questions that you have about the feature. It sounds like
you're creating quite a task list of work to do, but I don't understand
what your use case is so a lot of that work may be not be critical and you
I would really like to see the isolation feature in 0.94+, so I intend to
work with Jia and, if there is a successful result, support it going
forward in a manner like Stargate, even though I may not run it
personally (like I don't run Stargate). I take an expansive view of what
open source
+1
On 12/5/11 4:28 PM, Doug Meil doug.m...@explorysmedical.com wrote:
Hi folks-
Sorry to reopen this, but hopefully only briefly. I'm fine with renaming
it something else not book, but in taking a few samples of others would
anybody object to it being called Reference Guide? Reference Manual
I think it's good to remove the reflection when we can, more because it's
easier to catch compile-time errors than run-time. The perf is negligible
when you cache. As I recall, the problem here is the function was private
in older versions. We just need to make sure that we don't support
I agree with Todd's sentiments on unnecessarily coupling feature priority
with the availability of a patch. There's patches that we've developed
internally, then threw away because we couldn't tie it to a production use
case or thought a better design might be necessary. There's also patches
Committer questions. I accidentally checked in some changes to HBase
trunk with the Phabricator checkin format. I used svn propedit to correct
the log comments, so everything should look correct. However, this raised
a couple questions about committing code to HBase.
1. I know the required
To kick a dead horse, coupling the code diff with the CHANGES.txt
alteration is also a massive PITA because its much harder to run 'git
cherry-pick' across branches with it. It's far easier to directly port a
change from 92 to trunk to 89-fb if I don't have that pesky little file in
the way.
On
I've just created a new branch, 0.89. It has the current state of our largest
HBase deploy, Facebook Messaging. 0.89 was branched from HBase trunk back in
2010 before the release of HBase 0.90.0 so its pretty old. The current HBase
trunk has most of whats in 0.89 (and more) except for some
, Nicolas Spiegelberg
nspiegelb...@fb.com wrote:
I've just created a new branch, 0.89. It has the current state of our
largest HBase deploy, Facebook Messaging. 0.89 was branched from HBase
trunk back in 2010 before the release of HBase 0.90.0 so its pretty old.
The current HBase trunk has most
/28754491/
Focus is 0.92 stabilization. I don't think we can avoid wanting to talk about
0.94 features, though :P
Thanks,
Nicolas Spiegelberg
Chatted on IRC with Stack a couple others. We would like to host a Hackathon
+ Meetup on August 22 at Facebook in Palo Alto. All welcome. Unless
objection, I'll put up a posting on meetup.com. The hacking theme would be
HBase 0.92 stabilization. I'm sure 0.94 feature discussion will also
...@gmail.com wrote:
I was thinking about this too.
Is it possible to choose a day other than Monday (we have team meeting) ?
Thanks
On Thu, Aug 4, 2011 at 1:50 PM, Nicolas Spiegelberg
nspiegelb...@fb.comwrote:
Chatted on IRC with Stack a couple others. We would like to host a
Hackathon + Meetup
+1 for 0.92. This has received a lot of testing and code review internally. We
have already found/fixed a number of advanced bugs before even submitting the
initial patch. Very high quality.
On Jul 26, 2011, at 7:18 PM, Stack st...@duboce.net wrote:
Shall we commit the patch at
These slides were presented at FOSDEM, the conference where I just spoke about
FB HBase. I went to the talk.
There were some glaringly obvious problems that came out during the QA.
Someone asked how many regions he had. Didnt know. Are you network or IO or CPU
bottlenecked? Didn't know.
A great article for Bloom Filter rules of thumb:
http://corte.si/posts/code/bloom-filter-rules-of-thumb/
Note that only rules #1 #2 apply for our use case. Rule #3, while true, isn't
as big a worry because we use combinatorial generation for hashes, so the
number of 'expensive' hash
I don't think there's an explicit wiki. Which option depends on whether
your use case is calling get() for entire rows or for specific columns in
a row. It also depends on analyzing your workload to determine how likely
a row will be in every store file vs. a specific column. Also, since a
row
+1 for ELR.
I think having some data structure where we prepare the next stage of
sync() operations instead of holding the row lock over the sync would be a
big win for hot regions without a huge refactor. I think the other two
optimizations are useful to think about, but wouldn't have the same
Affects Versions: 0.21.0
Reporter: Nicolas Spiegelberg
Assignee: Nicolas Spiegelberg
Priority: Minor
Fix For: 0.21.0
Had a failure with testDynamicBloom on Hudson today. Will investigate, however
it would be nice to reproduce the problem to make sure
37 matches
Mail list logo