hive-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Nitin Vijayvargiya <nitinvija...@gmail.com>
Subject Re: Roaring Bitmap UDFs
Date Tue, 12 Dec 2017 04:25:22 GMT
Hi Prasanth,
Thanks, that was exactly what I was looking for. My main concern is speed, so I
tried going with the brickhouse implementation of HLL+, and ended up having to
make minor modifications to the code in order to have it run. My only concern is
that the precision check tests don't always pass. Given that I had to comment
out the@Ignore("not ready yet") over the test class, I'm unsure whether this
code is safe to use. Anyone else had this problem?
Thanks,Nitin  





On Mon, Dec 11, 2017 5:20 PM, Prasanth Jayachandran 
pjayachandran@hortonworks.com  wrote:
I did performance benchmark for roaring bitmaps when I added bloomfilters
(hyperloglog also shares the same bitset impl) to Orc and Hive. I found that
roaring bitmap is good at compression at the cost of speed. In a JMH benchmark,
observed around ~10x slowdown during insert and probe when using roaring bitmap
vs the raw bitset (long array without compression).  So it essentially boils
down to speed vs space tradeoff.  
Thanks  Prasanth

On Dec 11, 2017, at 3:58 PM, Nitin Vijayvargiya <nitinvijay94@gmail.com> wrote: 
 
Hi David,  
Thanks for the response. Yea, bloom filters are mostly for existential checks.
I'm looking for a way to preprocess data, and then perform operations like
union/intersection between them to find counts. Example: Number of distinct
users visiting website A over the last 5 days (union), intersected with the
number of distinct visitors visiting website B over the last 10 days (union).

Hyperloglog is the right tool for this, but if someone has done performance
benchmarking between HLL and Roaring BitMap, it would save me a lot of time.  
Thanks,  Nitin  





On Fri, Dec 8, 2017 7:08 PM, David Capwelldcapwell@gmail.comwrote:
Think bloom filter that's more dynamic.  It works well when cardinality is low,
but grows quickly to out cost bloom filter as cardinality grows.  
This data structure supports existence queries, but your email sounds like you
want count.  If so not really the best fit.

On Dec 8, 2017 5:00 PM, "Nitin Vijayvargiya" <nitinvijay94@gmail.com> wrote:
Hi all,  
I'm working on speeding up distinct count calculations, and it looks like
roaring bitmaps (RB) is the newest and meanest way for set operations. Anyone
here have experience with them? How was the performance compared to hyperloglog
and EWAH? A quick google search showed me that it's easier to find UDF
implementations of hyperloglog in presto and hive, but if the hype is real, it
might be worth spending the time to incorporate RB. Also, if anyone can point me
to reliable implementations of UDFs using RB, I would love to check it out and
test it myself =)  
Happy Holidays!  
Nitin
Mime
View raw message