Transcript a,2 - Stanford InfoLab

Towards a Synopsis
Warehouse
Peter J. Haas
IBM Almaden Research Center
San Jose, CA
1
Acknowledgements:
Kevin Beyer
Paul Brown
Rainer Gemulla (TU Dresden)
Wolfgang Lehner (TU Dresden)
Berthold Reinwald
Yannis Sismanis
2
Information Discovery for the Enterprise
Query: “Explain the product movement, buyer behavior, maximize the ROI on my product campaigns.”
Query: “The sales team is visiting company XYZ next week. What do they need to know about XYZ?”
Search
Business
Intelligence
Enterprise Repository
Data Analysis
&Similarity
Account
Content
Metadata
Order
Business objects
Customer
Business-Object
Discovery
Analyze, Integrate
Crawl, ETL
ERP (SAP), CRM, WBI
BPM, SCM
Structured
Company Data
ECM (reports, spreadsheets,
Financial docs (XBRL))
Semi-Structured
Syndicated Data Provider
Office documents
E-Mail, Product Manuals
Unstructured
Crawlable/deep Web
3
Motivation
• Challenge: Scalability
– Massive amounts of data at high speed
• Batches and/or streams
– Structured, semi-structured, unstructured data
• Want quick approximate analyses
–
–
–
–
Automated data integration and schema discovery
“Business object” identification
Quick approximate answers to queries
Data browsing/auditing
• Our approach: a warehouse of synopses
4
A Synopsis Warehouse
Synop.
Full-Scale
Warehouse Of
Data Partitions
Synop.
Synop.
S1,1
Warehouse
of Synopses
S1,2
Sn,m
merge
S*,*
S1-2,3-7
etc
5
Outline
• Synopsis 1: Uniform samples
– Background
– Creating and combining partitions
• Hybrid Bernoulli and Hybrid Reservoir algorithms
– Updating partitions
• Stable datasets: random pairing
• Growing datasets: resizing algorithms
• Synopsis 2: AKMV samples
– Goal: estimating the number of distinct values
– Our choice of synopsis and estimator
• Relation to other work
• Analysis based on theory of uniform order statistics
6
Synopsis 1: Uniform Samples
• Advantages
–
–
–
–
Flexible, used to produce other synopses
Mandatory if future use unknown a priori
Used in many statistical and mining methods
Building block for complex schemes
• Ex: stratified sampling
• Design goals
–
–
–
–
True uniformity (same sample size = same probability)
Bounded memory (no unpleasant surprises)
Memory efficiency (keep sample full)
Support for compressed samples
• Handle common case of few distinct values
• 80% of 1000 customer datasets had < 4000 distinct values
7
Classical Uniform Methods
• Bernoulli sampling
– Bern(q) independently includes each element with prob = q
– Random, uncontrollable sample size
• Binomially distributed, Nq on average
– Easy to merge: union of two Bern(q) samples is Bern(q)
• Reservoir sampling
– Creates uniform sample of fixed size k
– Insert first k elements into sample
– Then insert ith element with prob. pi = k / i
• replace random victim
• Vitter optimization: Simulate coin tosses
– Variants to handle large disk-resident samples
– Merging more complicated than Bernoulli
8
Drawback of Basic Methods
• Neither method is very compact
– Especially when few distinct values
• E.g., DS = (<A,500>,<B,300>)
• Stored as (A,A,…,A,B,B,…B) - 800 chars
• Concise sampling (GM 98)
– Compact: purge Bern(q) sample S if too large
• Bern(q’/q) subsample of S  Bern(q’) sample
– Not uniform (rare items under-represented)
9
New Sampling Methods (ICDE ’06)
• Two flavors:
– Hybrid reservoir (HR)
– Hybrid Bernoulli (HB)
• Properties
– Truly uniform
– Bounded footprint at all times
• Not as memory efficient as Concise Sampling
– Will store exact distribution if possible
– Samples stored in compressed form
– Merging algorithms available
10
Basic Ideas: Sample Creation
• Phase 1
– Start by storing 100% sample compactly
– Termination in Phase 1  exact distribution
• Abandon Phase 1 if footprint too big
– Take subsample and expand
– Fall back to reservoir(HR) or Bernoulli(HB) sampling (Phase 2)
– HB can revert to reservoir sampling (Phase 3)
• Compress sample upon termination
• If Phase 2 termination
– HR: uniform sample of size k
– HB: uniform, (almost) Bernoulli sample
• Stay within footprint at all times
– Messy details
11
Basic Ideas: Merging
• Both samples in Phase 2 (usual case)
– Bernoulli: equalize q’s and take union
• Take subsample to equalize q’s
– Reservoir: take subsamples and merge
• Random (hypergeometric) subsample size
• Corner cases
– One sample in Phase 1, etc.
– See ICDE ’06 paper for details
12
HB versus HR
• Advantages:
– HB samples are cheaper to merge
• Disadvantages:
– HR sampling controls sample size better
– Need to know partition size in advance
• For subsampling during sample creation
– Engineering approximation required
13
Speedup: HB Sampling
You derive “speed-up” advantages from
parallelism with up to about 100 partitions.
14
Speedup: HR Sampling
Similar results to previous slide, but merging HR
samples is more complex than HB samples.
15
Linear Scale-Up
HB Sampling
HR Sampling
16
Updates Within a Partition
• Arbitrary inserts/deletes (updates trivial)
• Previous goals still hold
– True uniformity
– Bounded sample size
– Keep sample size close to upper bound
• Also: minimize/avoid base-data access
17
New Algorithms (VLDB ’06+)
• Stable datasets: Random pairing
– Generalizes reservoir/stream sampling
• Handles deletions
• Avoids base-data accesses
– Dataset insertions paired randomly with “uncompensated deletions”
• Only requires counters (cg, cb) of “good” and “bad” UD’s
• Insert into sample with probability cb / (cb + cg)
– Extended sample-merging algorithm
• Growing datasets: Resizing
– Theorem: can’t avoid base-data access
– Main ideas:
• Temporarily convert to Bern(q): requires base-data access
• Drift up to new size (stay within new footprint at all times)
• Choose q optimally to reduce overall resizing time
– Approximate and Monte Carlo methods
18
Multiset Sampling (PODS ’07)
• Bernoulli samples over multisets (w. deletions)
– When boundedness not an issue
– Problem: how to handle deletions (pairing?)
• Idea: maintain “tracking counter”
– # inserts into DS since first insertion into sample (GM98)
• Can exploit tracking counter
– To estimate frequencies, sums, avgs
• Unbiased (except avg) and low variance
– To estimate # distinct values
• Maintaining tracking counter
– Subsampling: new algorithm
– Merging: negative result
– Synopsis-warehouse subsampling and merging still OK
19
Synopsis 2:
AKMV Synopses (SIGMOD ’07)
• Goal: Estimate # distinct values
– Dataset similarity (Jaccard distance)
– Key detection
– Data cleansing
• Within warehouse framework
– Must handle multiset union, intersection,
difference
20
KMV Synopsis
• Used for a base partition
• Synopsis: k smallest hashed values
– vs bitmaps (e.g., logarithmic counting)
• Need inclusion/exclusion to handle intersection
• Less accuracy, poor scaling
– vs sample counting
• Random size K (between k/2 and k)
– vs Bellman [DJMS02]
• minHash for k independent hash functions
• O(k) time per arriving value, vs O(log k)
• Can view as uniform sample of DV’s
21
The Basic Estimator
• Estimator: Dˆ  (k  1) / U( k )
– U(k) = kth smallest hashed value
• Properties (theory of uniform order statistics)
– Normalized hashed values “look like” i.i.d. uniform[0,1] RVs
D(D  1) (D  r  1)
E [Dˆ ]  D and E [Dˆ r ]  (k  1)r
(k  1)(k  2) (k  r )
E | Dˆ  D | / D  is bounded in D (can also get confidence bounds)
Asymptotically efficient as k   (minimal variance)
• Large-D scenario (simpler formulas)
– Theorem: U(k) approx.= sum of k i.i.d exp(D) random variables
– Analysis coincides with [Cohen97]
– Can use simpler formulas to choose synopsis size
22
Compound Partitions
• Given a multiset expression E
– In terms of base partitions A1,…,An
– Union, intersection, multiset difference
• Augmented KMV synopsis
– Augment with counters
– AKMV synopses are closed under multiset operations
• Modified unbiased estimator for # DVs in E
• See SIGMOD ’07 for details
23
Experimental Comparison
Absolute Relative Error
0.1
0.08
0.06
0.04
0.02
0
Unbiased
SDLogLog Sample-Counting Unbiased-baseline
24
For More Details
•
"Toward automated large scale information integration and discovery."
P. Brown, P. J. Haas, J. Myllymaki, H. Pirahesh, B. Reinwald, and Y.
Sismanis. In Data Management in a Connected World, T. Härder and
W. Lehner, eds. Springer-Verlag, 2005.
•
“Techniques for warehousing of sample data”. P. G. Brown and P. J.
Haas. ICDE ‘06.
•
“A dip in the reservoir: maintaining sample synopses of evolving
datasets”. R. Gemulla, W. Lehner, and P. J. Haas. VLDB ‘06.
•
“Maintaining Bernoulli samples over evolving multisets”.
R. Gemulla, W. Lehner, and P. J. Haas. PODS ‘07.
•
“On Synopses for DistinctValue Estimation Under Multiset Operations”
K. Beyer, P. J. Haas, B. Reinwald, Y. Sismanis, and R. Gemulla.
SIGMOD 2007.
25
Backup Slides
26
Bernoulli Sampling
– Bern(q) independently includes each element with probability q
– Random, uncontrollable sample size
– Easy to merge Bernoulli samples: union of 2 Bern(q) samp’s = Bern(q)
q = 1/3
2 /3
1 /3
+t1
1
2 /3
+t2
2 /3 1 /3
+t3
3
30%
2 /3
1 /3
1 2
2
1
2 /3 1 /3
2 /3 1 /3
2
15% 15%
1 /3
2 /3
1 /3
2 3
1
1 3
1 2
1 2 3
7%
15%
7%
7%
4%
27
Reservoir Sampling (Example)
• Sample size M = 2
+t1 +t2
1 2
1 /3
+t3
+ t4
1 /3
1 /3
100%
1 2
3 2
2 /4
33%
1/4
1 /4
1 2
4 2
16%
8%
1 3
2/4
33%
1 /4
1/4
2/4
33%
1 /4
1/4
1 4
3 2
4 2
3 4
1 3
4 3
1 4
8%
16%
8%
8%
16%
8%
8%
28
Concise-Sampling Example
•
Dataset
– D = { a, a, a, b, b, b }
•
Footprint
– F = one <value, #> pair
•
Three (possible) samples of size = 3
– S1 = { a, a, a }, S2 = { b, b, b }, S3 = { a, a, b }.
– S1 = {<a,3>}, S2 = {<b,3>}, S3 = {<a,2>,<b,1>}.
•
Three samples should have with equal likelihood
– But Prob(S1) = Prob(S2) > 0 and Prob(S3) = 0
•
In general:
– Concise sampling under-represents ‘rare’ population elements
29
Hybrid Reservoir (HR) Sampling
Ex: Sample capacity = two <v,#> pairs or three values
Phase 1 (Maintain exact frequency distribution)
+a +a
{<a,2>}
+a
{<a,3>}
+b
{<a,3>,<b,1>}
{<a,3>,b}
+b
{<a,3>,<b,2>}
+c
Phase 2 (Reservoir sampling)
{a,<b,2>}
(subsample)
{a,b,b}
(expand)
{c,b,b}
+d
{c,b,d}
…
…
+a
{a,a,a}
done
{<a,3>}
(reservoir sampling)
(compress)
30
Subsampling in HB Algorithm
• Goal: find q such that P{|S| > nF} = p
• Solve numerically:

|D|
  q (1 q)
|D|
j  nF 1 j
|D| j
j
p
• Approximate solution (< 3% error):
q



| D | 2nF  zp2  zp | D | | D | zp2  4 | D | nF  4nF2

2 | D | | D |  zp2


31
Merging HB Samples
•
If both samples in Phase 2 (the usual case)
– Choose q as before (w.r.t. |D1 U D2|)
– Convert both samples to compressed Bern(q)
[Use Bern(q’/q) trick as in Concise Sampling]
– If union of compressed samples fits in memory
then join and exit
else use reservoir sampling (unlikely)
•
See paper for remaining cases
32
Merging a Pair of HR Samples
• If at least one sample in Phase 1
– Treat Phase-1 sample as “data stream”
– Run HR sampling algorithm
• If both samples in Phase 2
– Set k = min(|S1|, |S2|)
– Select L elements from S1 and k – L from S2
• L has hypergeometric distribution on {0,1,…,k}
– Distribution depends on |D1|, |D2|
• Take (compressed) reservoir subsamples of S1, S2
• Join (compressed union) and exit
33
Generating Realizations of L
L is a random variable with probability mass function
P(l) = P{ L=l } given by:
 | D1 |   | D2 | 

 l 
 
k l 





P (l ) 
 | D1 |  | D2 | 




k


•
Simplest implementation
–
–
•
for l = 0, 1, …. k-1
Compute P recursively
Use inversion method (probe cumulative distribution at each merge)
Optimizations when |D|’s and |S|’s unchanging
–
Use alias methods to generate L from cached distributions in O(1) time
34
Naïve/Prior Approaches
Algorithm
Technique
(RS with deletions)
conduct deletions, continue with
smaller sample
Naïve
use insertions to immediately refill
the sample
Comments
unstable
not uniform
RS with resampling let sample size decrease, but
occasionally recompute
expensive, unstable
CAR(WOR)
immediately sample from base data
to refill the sample
stable but expensive
Bernoulli sampling
with purging
“coin flip” sampling with deletions,
purge if too large
inexpensive but unstable
Passive sampling
developed for data streams (sliding
windows only)
special case of our RP
algorithm
Distinct-value
sampling
tailored for multiset populations
expensive, low space
efficiency in our setting
Counting samples
Modification of concise sampling
Not uniform
35
Random Pairing
+t1 +t2
1 2
1/3
+t3
1 2
-t2
-t3
1
1
+t5
3 2
1
1
1
3
1
1
1 3
1
1 3
1
1
1/2
+t4
1/3
1/3
1
1/2
1 4
1
1/2
1/2
4
4
1
1
1/2
1 4
1
1/2
1
1
1 5
1 4
4 5
4 5
1 4
1 5
16%
16%
16%
16%
16%
16%
36
Performance
100
100K sample
10 million operations
RP
CAR
CARW OR
RSR
BSP
Cost
10
1
0.1
0
2
4
6
8
10
Sample Size (%)
Dataset Size (millions)
100
90
80
RP
BSP
70
0
1
2
3
4
Number of Operations (millions)
5
37
A Negative Result
• Theorem
– Any resizing algorithm MUST access base data
• Example
– data set
1 2 3 4
– samples of size 2
– new data set
– samples of size 3
1 2
1 3
1 4
2 3
2 4
3 4
16%
16%
16%
16%
16%
16%
1 2 3 4 5 6 ...
1 2 3
1 2 5
0%
>0%
38
Resizing: Phase 1
Conversion to Bernoulli sample
– Given q, randomly determine sample size
• U = Binomial(|D|,q)
– Reuse S to create Bernoulli sample
• Subsample if U < |S|
• Else sample additional tuples (base data access)
– Choice of q
• small  less base data accesses
• large  more base data accesses
39
Resizing: Phase 2
Run Bernoulli sampling
– Include new tuples with probability q
– Delete from sample as necessary
– Eventually reach new sample size
– Revert to reservoir sampling
– Choice of q
• small  long drift time
• large  short drift time
40
Choosing q (Inserts Only)
• Expected Phase 1 (conversion) time


| D | M

E [T1]  ta | D | ln 

 | D | M  min(| D | q, M ')  M  


• Expected Phase 2 (drifting) time
t b M '  | D | q 

E [T2 ] 
q
• Choose q to minimize E[T1] + E[T2]
41
Resizing Behavior
• Example (dependence on base-access cost):
– resize by 30% if sampling fraction drops below 9%
– dependent on costs of accessing base data
Low costs
immediate resizing
Moderate costs
High costs
combined solution
degenerates to Bernoulli
sampling
42
Choosing q (w. Deletes)
• Simple approach (insert prob. = p > 0.5)
– Expected change in partition size (Phase 2)
• (p)(1)+(1-p)(-1) = 2p-1
– So scale Phase 2 cost by 1/(2p-1)
• More sophisticated approach
– Hitting time of Markov chain to boundary
– Stochastic approximation algorithm
• Modified Kiefer-Wolfowitz
43
Estimating the DV Count
• Exact computation via sorting
– Usually infeasible
• Sampling-based estimation
– Very hard problem (need large samples)
• Probabilistic counting schemes
– Single-pass, bounded memory
– Several flavors (mostly bit-vector synopses)
• Linear counting (ASW87)
• Logarithmic counting (FM85,WVT90,AMS, DF03)
• Sample counting (ASW87,Gi01, BJKST02)
44
Intuition
• Look at spacings
– Example with k = 4 and D = 7:
V1 V2
0
–
–
–
–
–
x x
V3
V4
x x
x x
x
1
E[V]  1 / D so that D  1 / E[V]
Estimate D as 1 / Avg(V1,…,Vk)
I.e., as k / Sum(V1,…,Vk)
I.e., as k / u(k)
Upward bias (Jensen’s inequality) so change k to k-1
45