273 lines
29 KiB
TeX
273 lines
29 KiB
TeX
%!TEX root=./main.tex
|
|
%root: main.tex
|
|
\section{Introduction}\label{sec:intro}
|
|
|
|
This work explores the problem of computing the expectation of the multiplicity of a tuple in the result of a query over a \abbrCTIDB, a type of probabilistic database with bag semantics where the multiplicity of a tuple is a random variable with range $[0,\bound]$ for some fixed constant $\bound$ and multiplicities assigned to any two tuples are independent of each other.
|
|
Formally, a \abbrCTIDB,
|
|
$\pdb = \inparen{\worlds, \bpd}$ consists of a set of tuples $\tupset$ and a probability distribution $\bpd$ over all possible worlds generated by assigning each tuple $\tup \in \tupset$ a multiplicity in the range $[0,\bound]$.
|
|
Any such world can be encoded as a vector from $\worlds$, the set of all vectors of length $\numvar=\abs{\tupset}$ such that each index corresponds to a distinct $\tup \in \tupset$ storing its multiplicity.
|
|
A given world $\worldvec \in\worlds$ can be interpreted as follows: for each $\tup \in \tupset$, $\worldvec_{\tup}$ is the multiplicity of $\tup$ in $\worldvec$. Given that the multiplicities of tuples are independent events, the probability distribution $\bpd$ can be expressed compactly by assigning each tuple a (disjoint) probability distribution over $[0,\bound]$. Let $\prob_{\tup,j}$ denote the probability that tuple $\tup$ is assigned multiplicity $j$. The probability of a particular world $\worldvec$ is then $\prod_{\tup \in \tupset} \prob_{\tup,\worldvec_{\tup}}$.
|
|
|
|
Allowing for $\leq \bound$ multiplicities across all tuples gives rise to having $\leq \inparen{\bound+1}^\numvar$ possible worlds instead of the usual $2^\numvar$ possible worlds of a $1$-\abbrTIDB, which (assuming set query semantics), is the same as the traditional set \abbrTIDB.
|
|
In this work, since we are generally considering bag query input, we will only be considering bag query semantics. We denote by $\query\inparen{\worldvec}\inparen{\tup}$ the multiplicity of $\tup$ in query $\query$ over possible world $\worldvec\in\worlds$.
|
|
|
|
We can formally state our problem of computing the expected multiplicity of a result tuple as:
|
|
|
|
\begin{Problem}\label{prob:expect-mult}
|
|
Given a \abbrCTIDB $\pdb = \inparen{\worlds, \bpd}$, $\raPlus$ query\footnote{
|
|
An $\raPlus$ query is a query expressed in positive relational algebra, i.e., using only the relational algebra operators selection ($\select$), projection ($\project$), natural join ($\join$) and union ($\union$).
|
|
}
|
|
$\query$, and result tuple $\tup$, compute the expected multiplicity of $\tup$: $\expct_{\rvworld\sim\bpd}\pbox{\query\inparen{\rvworld}\inparen{\tup}}$.
|
|
\end{Problem}
|
|
|
|
\begin{figure}[t!]
|
|
\begin{align*}
|
|
&\begin{aligned}[t]
|
|
&\polyqdt{\project_A(\query)}{\gentupset}{\tup} =\\
|
|
&~~\sum_{\tup': \project_A(\tup') = \tup} \polyqdt{\query}{\gentupset}{\tup'}
|
|
\end{aligned}
|
|
&
|
|
&\begin{aligned}[t]
|
|
&\polyqdt{\query_1 \union \query_2}{\gentupset}{\tup} =\\
|
|
&\qquad \polyqdt{\query_1}{\gentupset}{\tup} + \polyqdt{\query_2}{\gentupset}{\tup}\\
|
|
\end{aligned}\\
|
|
&\begin{aligned}
|
|
&\polyqdt{\select_\theta(\query)}{\gentupset}{\tup} =\\
|
|
&~~ \begin{cases}
|
|
\polyqdt{\query}{\gentupset}{\tup} & \text{if }\theta(\tup) \\
|
|
0 & \text{otherwise}.
|
|
\end{cases}
|
|
\end{aligned}
|
|
&
|
|
&\begin{aligned}
|
|
&\polyqdt{\query_1 \join \query_2}{\gentupset}{\tup} =\\
|
|
&\qquad\polyqdt{\query_1}{\gentupset}{\project_{\attr{\query_1}}{\tup}}\\
|
|
&\qquad\cdot\polyqdt{\query_2}{\gentupset}{\project_{\attr{\query_2}}{\tup}}
|
|
\end{aligned}\\
|
|
&&&\polyqdt{\rel}{\gentupset}{\tup} = X_\tup
|
|
\end{align*}%\\[-10mm]
|
|
\setlength{\abovecaptionskip}{-0.25cm}
|
|
\caption{Construction of the lineage (polynomial) for an $\raPlus$ query $\query$ over an arbitrary deterministic database $\gentupset$, where $\vct{X}$ consists of all $X_\tup$ over all $\rel$ in $\gentupset$ and $\tup$ in $\rel$. Here $\gentupset.\rel$ denotes the instance of relation $\rel$ in $\gentupset$. Please note, after we introduce the reduction to $1$-\abbrBIDB, the base case will be expressed alternatively.}
|
|
\label{fig:nxDBSemantics}
|
|
\vspace{-0.53cm}
|
|
\end{figure}
|
|
|
|
It is natural to explore computing the expected multiplicity of a result tuple as this is the analog for computing the marginal probability of a tuple in a set \abbrPDB.
|
|
In this work we will assume that $c =\bigO{1}$ since this is what is typically seen in practice.
|
|
Allowing for unbounded $c$ is an interesting open problem.
|
|
|
|
\mypar{Hardness of Set Query Semantics and Bag Query Semantics}
|
|
Set query evaluation semantics over $1$-\abbrTIDB\xplural have been studied extensively, and the data complexity of the problem in general has been shown by Dalvi and Suicu to be \sharpphard\cite{10.1145/1265530.1265571}. For our setting, there exists a trivial polytime algorithm to compute~\Cref{prob:expect-mult} for any $\raPlus$ query over a \abbrCTIDB due to linearity of expection (see~\Cref{sec:intro-poly-equiv}).
|
|
Since we can compute~\Cref{prob:expect-mult} in polynomial time, the interesting question that we explore deals with analyzing the hardness of computing expectation using fine-grained analysis and parameterized complexity, where we are interested in the exponent of polynomial runtime.
|
|
|
|
Specifically, in this work we ask if~\Cref{prob:expect-mult} can be solved in time linear in the runtime of an analogous deterministic query which we make more precise shortly.
|
|
If this is true, then this would open up the way for deployment of \abbrCTIDB\xplural in practice. To analyze this question we denote by $\timeOf{}^*(Q,\pdb)$ the optimal runtime complexity of computing~\Cref{prob:expect-mult} over \abbrCTIDB $\pdb$.
|
|
|
|
Let $\qruntime{\query,\gentupset,\bound}$ (see~\Cref{sec:gen} for further details) denote the runtime for query $\query$, deterministic database $\gentupset$, and multiplicity bound $\bound$. This paper considers $\raPlus$ queries for which order of operations is \emph{explicit}, as opposed to other query languages, e.g. Datalog, UCQ. Thus, since order of operations affects runtime, we denote the optimized $\raPlus$ query picked by an arbitrary production system as $\optquery{\query} = \min_{\query'\in\raPlus, \query'\equiv\query}\qruntime{\query', \gentupset, \bound}$. Then $\qruntime{\optquery{\query}, \gentupset,\bound}$ is the runtime for the optimized query.\footnote{Note that our work applies to any $\query \in\raPlus$, which implies that specific heuristics for choosing an optimized query can be abstracted away, i.e., our work does not consider heuristic techniques.}
|
|
|
|
\begin{table*}[t!]
|
|
\centering
|
|
\begin{tabular}{|p{0.43\textwidth}|p{0.12\textwidth}|p{0.35\textwidth}|}
|
|
\hline
|
|
\textbf{Lower bound on $\timeOf{}^*(\qhard,\pdb)$} & \textbf{Num.} $\bpd$s
|
|
& \textbf{Hardness Assumption}\\
|
|
\hline
|
|
$\Omega\inparen{\inparen{\qruntime{\optquery{\qhard}, \tupset, \bound}}^{1+\eps_0}}$ for {\em some} $\eps_0>0$ & Single & Triangle Detection hypothesis\\
|
|
$\omega\inparen{\inparen{\qruntime{\optquery{\qhard}, \tupset, \bound}}^{C_0}}$ for {\em all} $C_0>0$ & Multiple &$\sharpwzero\ne\sharpwone$\\
|
|
$\Omega\inparen{\inparen{\qruntime{\optquery{\qhard}, \tupset, \bound}}^{c_0\cdot k}}$ for {\em some} $c_0>0$ & Multiple & \Cref{conj:known-algo-kmatch}\\
|
|
\hline
|
|
\end{tabular}
|
|
\caption{Our lower bounds for a specific hard query $\qhard$ parameterized by $k$. For $\pdb = \inset{\worlds, \bpd}$ those with `Multiple' in the second column need the algorithm to be able to handle multiple $\bpd$, i.e. probability distributions (for a given $\tupset$). The last column states the hardness assumptions that imply the lower bounds in the first column ($\eps_o,C_0,c_0$ are constants that are independent of $k$).}
|
|
\label{tab:lbs}
|
|
\vspace{-0.73cm}
|
|
\end{table*}
|
|
|
|
\mypar{Our lower bound results}
|
|
Our question is whether or not it is always true that $\timeOf{}^*\inparen{\query, \pdb}\leq\qruntime{\optquery{\query}, \tupset, \bound}$. Unfortunately this is not the case.
|
|
~\Cref{tab:lbs} shows our results.
|
|
|
|
Specifically, depending on what hardness result/conjecture we assume, we get various weaker or stronger versions of {\em no} as an answer to our question. To make some sense of the other lower bounds in \Cref{tab:lbs}, we note that it is not too hard to show that $\timeOf{}^*(Q,\pdb) \le \bigO{\inparen{\qruntime{\optquery{\query}, \tupset, \bound}}^k}$, where $k$ is the join width (our notion of join width follows from~\Cref{def:degree-of-poly} and~\Cref{fig:nxDBSemantics}.) of the query $\query$ over all result tuples $\tup$ (and the parameter that defines our family of hard queries).
|
|
|
|
What our lower bound in the third row says is that one cannot get more than a polynomial improvement over essentially the trivial algorithm for~\Cref{prob:expect-mult}.
|
|
However, this result assumes a hardness conjecture that is not as well studied as those in the first two rows of the table (see \Cref{sec:hard} for more discussion on the hardness assumptions). Further, we note that existing results\footnote{This claim follows from known results for the problem of counting $k$-cliques, where for a query $\query$ over database $\tupset$ that counts the number of $k$-cliques. Specifically, a lower bound of the form $\Omega\inparen{n^{1+\eps_0}}$ for {\em some} $\eps_0>0$ follows from the triangle detection hypothesis (this like our result is for $k=3$). Second, a lower bound of $\omega\inparen{n^{C_0}}$ for {\em all} $C_0>0$ under the assumption $\sharpwzero\ne\sharpwone$ for counting $k$-clique~\cite{10.5555/645413.652181}. Finally, a lower bound of $\Omega\inparen{n^{c_0\cdot k}}$ for {\em some} $c_0>0$ was shown by~\cite{CHEN20061346} (under the strong exponential time hypothesis).
|
|
} already imply the claimed lower bounds if we were to replace the $\qruntime{\optquery{\query}, \tupset, \bound}$ by just $\numvar$ (indeed these results follow from known lower bounds for deterministic query processing). Our contribution is to then identify a family of hard queries where deterministic query processing is `easy' but computing the expected multiplicities is hard.
|
|
|
|
\mypar{Our upper bound results} We introduce a $(1\pm \epsilon)$-approximation algorithm that computes ~\Cref{prob:expect-mult} in time $O_\epsilon\inparen{\qruntime{\optquery{\query}, \tupset, \bound}}$. This means, when we are okay with approximation, that we solve~\Cref{prob:expect-mult} in time linear in the size of the deterministic query and bag \abbrPDB\xplural are deployable in practice.
|
|
In contrast, known approximation techniques (\cite{DBLP:conf/icde/OlteanuHK10,DBLP:journals/jal/KarpLM89}) in set-\abbrPDB\xplural need time $\Omega(\qruntime{\optquery{\query}, \tupset, \bound}^{2k})$
|
|
(see \Cref{sec:karp-luby}).
|
|
Further, our approximation algorithm works for a more general notion of bag \abbrPDB\xplural beyond \abbrCTIDB\xplural
|
|
(see \Cref{subsec:tidbs-and-bidbs}).
|
|
|
|
\subsection{Polynomial Equivalence}\label{sec:intro-poly-equiv}
|
|
A common encoding of probabilistic databases (e.g., in \cite{IL84a,Imielinski1989IncompleteII,Antova_fastand,DBLP:conf/vldb/AgrawalBSHNSW06} and many others) relies on annotating tuples with lineages or propositional formulas that describe the set of possible worlds that the tuple appears in. The bag semantics analog is a provenance/lineage polynomial (see~\Cref{fig:nxDBSemantics}) $\apolyqdt$~\cite{DBLP:conf/pods/GreenKT07}, a polynomial with non-zero integer coefficients and exponents, over variables $\vct{X}$ encoding input tuple multiplicities. Evaluating a lineage polynomial for a query result tuple $t_{out}$ by, for each tuple $\tup_{in}$, assigning the variable $X_{t_{in}}$ encoding the tuple's multiplicity to the tuple's multiplicity in the possible world yields the multiplicity of the $\tup_{out}$ in the query result for this world.
|
|
|
|
We drop $\query$, $\tupset$, and $\tup$ from $\apolyqdt$ when they are clear from the context or irrelevant to the discussion. We now specify the problem of computing the expectation of tuple multiplicity in the language of lineage polynomials:
|
|
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
|
\begin{Problem}[Expected Multiplicity of Lineage Polynomials]\label{prob:bag-pdb-poly-expected}
|
|
Given an $\raPlus$ query $\query$, \abbrCTIDB $\pdb$ and result tuple $\tup$, compute the expected
|
|
multiplicity of the polynomial $\apolyqdt$ (i.e., $\expct_{\vct{W}\sim \pdassign}\pbox{\apolyqdt(\vct{W})}$, where $\vct{W} \in \worlds$).
|
|
\end{Problem}
|
|
We note that computing \Cref{prob:expect-mult}
|
|
is equivalent (yields the same result as) to computing \Cref{prob:bag-pdb-poly-expected} (see \Cref{prop:expection-of-polynom}).
|
|
|
|
All of our results rely on working with a {\em reduced} form $\inparen{\rpoly}$ of the lineage polynomial $\poly$. In fact, it turns out that for the $1$-\abbrTIDB case, computing the expected multiplicity (over bag query semantics) is {\em exactly} the same as evaluating this reduced polynomial over the probabilities that define the $1$-\abbrTIDB. This is also true when the query input(s) is a block independent disjoint probabilistic database~\cite{DBLP:conf/icde/OlteanuHK10} (bag query semantics with tuple multiplicity at most $1$), for which the proof of~\Cref{lem:tidb-reduce-poly} (introduced shortly) holds .
|
|
|
|
Next, we motivate this reduced polynomial.
|
|
Consider the query $\query_1$ defined as follows over the bag relations of \Cref{fig:two-step}:
|
|
|
|
\begin{lstlisting}
|
|
SELECT DISTINCT 1 FROM T $t_1$, R r, T $t_2$
|
|
WHERE $t_1$.Point = r.Point$_1$ AND $t_2$.Point = r.Point$_2$
|
|
\end{lstlisting}
|
|
|
|
It can be verified that $\poly\inparen{A, B, C, E, X, Y, Z}$ for the sole result tuple of $\query_1$ is $AXB + BYE + BZC$. Now consider the product query $\query_1^2 = \query_1 \times \query_1$.
|
|
The lineage polynomial for $Q_1^2$ is given by $\poly_1^2\inparen{A, B, C, E, X, Y, Z}$
|
|
$$
|
|
=A^2X^2B^2 + B^2Y^2E^2 + B^2Z^2C^2 + 2AXB^2YE + 2AXB^2ZC + 2B^2YEZC.
|
|
$$
|
|
To compute $\expct\pbox{\poly_1^2}$ we can use linearity of expectation and push the expectation through each summand. To keep things simple, let us focus on the monomial $\poly_1^{\inparen{ABX}^2} = A^2X^2B^2$ as the procedure is the same for all other monomials of $\poly_1^2$. Let $\randWorld_X$ be the random variable corresponding to a lineage variable $X$. Because the distinct variables in the product are independent, we can push expectation through them yielding $\expct\pbox{\randWorld_A^2\randWorld_X^2\randWorld_B^2}=\expct\pbox{\randWorld_A^2}\expct\pbox{\randWorld_X^2}\expct\pbox{\randWorld_B^2}$. Since $\randWorld_A, \randWorld_B\in \inset{0, 1}$ we can further derive $\expct\pbox{\randWorld_A}\expct\pbox{\randWorld_X^2}\expct\pbox{\randWorld_B}$ by the fact that for any $W\in \inset{0, 1}$, $W^2 = W$. Observe that if $X\in\inset{0, 1}$, then we further would have $\expct\pbox{\randWorld_A}\expct\pbox{\randWorld_X}\expct\pbox{\randWorld_B} = \prob_A\cdot\prob_X\cdot\prob_B$ (denoting $\probOf\pbox{\randWorld_A = 1} = \prob_A$) $= \rpoly_1^{\inparen{ABX}^2}\inparen{\prob_A, \prob_X, \prob_B}$ (see $ii)$ of~\Cref{def:reduced-poly}). However, in this example, we get stuck with $\expct\pbox{\randWorld_X^2}$, since $\randWorld_X\in\inset{0, 1, 2}$ and for $\randWorld_X \gets 2$, $\randWorld_X^2 \neq \randWorld_X$.
|
|
|
|
Denote the variables of $\poly$ to be $\vars{\poly}.$ In the \abbrCTIDB setting, $\poly\inparen{\vct{X}}$ has an equivalent reformulation $\inparen{\refpoly{}\inparen{\vct{X_R}}}$ that is of use to us, where $\abs{\vct{X_R}} = \bound\cdot\abs{\vct{X}}$ . Given $X_\tup \in\vars{\poly}$ and integer valuation $X_\tup \in\inset{0,\ldots, c}$. We can replace $X_\tup$ by $\sum_{j\in\pbox{\bound}}jX_{\tup, j}$ where the variables $\inparen{X_{\tup, j}}_{j\in\pbox{\bound}}$ are disjoint with integer assignments $X_\tup\in\inset{0, 1}$. Then for any $\worldvec\in\worlds$ and corresponding reformulated world $\worldvec_{\vct{R}}\in\inset{0, 1}^{\tupset\bound}$, we set $\worldvec_{\vct{R}_{\tup, j}} = 1$ for $\worldvec_\tup = j$, while $\worldvec_{\vct{R}_{\tup, j'}} = 0$ for all $j'\neq j\in\pbox{\bound}$. By construction then $\poly\inparen{\vct{X}}\equiv\refpoly{}\inparen{\vct{X_R}}$ $\inparen{\vct{X_R} = \vars{\refpoly{}}}$ since for any integer valuation $X_\tup\in\pbox{\bound}$, $X_j\in\inset{0, 1}$ we have the equality $X_\tup = j = \sum_{j\in\pbox{\bound}}jX_j$.
|
|
|
|
Considering again our example,
|
|
\begin{multline*}
|
|
\refpoly{1, }^{\inparen{ABX}^2}\inparen{A, X, B} = \poly_1^{\inparen{AXB}^2}\inparen{\sum_{j_1\in\pbox{\bound}}j_1A_{j_1}, \sum_{j_2\in\pbox{\bound}}j_2X_{j_2}, \sum_{j_3\in\pbox{\bound}}j_3B_{j_3}} \\
|
|
= \inparen{\sum_{j_1\in\pbox{\bound}}j_1A_{j_1}}^2\inparen{\sum_{j_2\in\pbox{\bound}}j_2X_{j_2}}^2\inparen{\sum_{j_3\in\pbox{\bound}}j_3B_{j_3}}^2.
|
|
\end{multline*}
|
|
Since the set of multiplicities for tuple $\tup$ by nature are disjoint we can drop all cross terms and have $\refpoly{1, }^2 = \sum_{j_1, j_2, j_3 \in \pbox{\bound}}j_1^2A^2_{j_1}j_2^2X_{j_2}^2j_3^2B^2_{j_3}$. Computing expectation we get $\expct\pbox{\refpoly{1, }^2}=\sum_{j_1,j_2,j_3\in\pbox{\bound}}j_1^2j_2^2j_3^2\expct\pbox{\randWorld_{A_{j_1}}}\expct\pbox{\randWorld_{X_{j_2}}}\expct\pbox{\randWorld_{B_{j_3}}}$, since we now have that all $\randWorld_{X_j}\in\inset{0, 1}$.
|
|
This leads us to consider a structure related to the lineage polynomial.
|
|
|
|
\begin{Definition}\label{def:reduced-poly}
|
|
For any polynomial $\poly\inparen{\inparen{X_\tup}_{\tup\in\tupset}}$ define the reformulated polynomial $\refpoly{}\inparen{\inparen{X_{\tup, j}}_{\tup\in\tupset, j\in\pbox{\bound}}}
|
|
$ to be the polynomial $\refpoly{}$ = $\poly\inparen{\inparen{\sum_{j\in\pbox{\bound}}j\cdot X_{\tup, j}}_{\tup\in\tupset}}
|
|
$ and ii) define the \emph{reduced polynomial} $\rpoly\inparen{\inparen{X_{\tup, j}}_{\tup\in\tupset, j\in\pbox{\bound}}}
|
|
$ to be the polynomial resulting from converting $\refpoly{}$ into the standard monomial basis (\abbrSMB),
|
|
\footnote{
|
|
This is the representation, typically used in set-\abbrPDB\xplural, where the polynomial is reresented as sum of `pure' products. See \Cref{def:smb} for a formal definition.
|
|
}
|
|
removing all monomials containing the term $X_{\tup, j}X_{\tup, j'}$ for $\tup\in\tupset, j\neq j'\in\pbox{c}$, and setting all \emph{variable} exponents $e > 1$ to $1$.
|
|
\end{Definition}
|
|
Continuing with the example
|
|
\footnote{
|
|
To save clutter we do not show the full expansion for variables with greatest multiplicity $= 1$ since e.g. for variable $A$, the sum of products itself evaluates to $1^2\cdot A^2 = A$.
|
|
}
|
|
$\poly_1^2\inparen{A, B, C, E, X_1, X_2, Y, Z}$ we have
|
|
\begin{multline*}
|
|
\rpoly_1^2(A, B, C, E, X_1, X_2, Y, Z) = \\
|
|
A\inparen{\sum\limits_{j\in\pbox{\bound}}j^2X_j}B + BYE + BZC + 2A\inparen{\sum\limits_{j\in\pbox{\bound}}j^2X_j}BYE + 2A\inparen{\sum\limits_{j\in\pbox{\bound}}j^2X_j}BZC + 2BYEZC =\\
|
|
ABX_1 + AB\inparen{2}^2X_2 + BYE + BZC + 2AX_1BYE + 2A\inparen{2}^2X_2BYE + 2AX_1BZC + 2A\inparen{2}^2X_2BZC + 2BYEZC.
|
|
\end{multline*}
|
|
Note that we have argued that for our specific example the expectation that we want is $\rpoly_1^2(\probOf\inparen{A=1},$ $\probOf\inparen{B=1}, \probOf\inparen{C=1}), \probOf\inparen{E=1}, \probOf\inparen{X_1=1}, \probOf\inparen{X_2=1}, \probOf\inparen{Y=1}, \probOf\inparen{Z=1})$.
|
|
\Cref{lem:tidb-reduce-poly} generalizes the equivalence to {\em all} $\raPlus$ queries on \abbrCTIDB\xplural (proof in \Cref{subsec:proof-exp-poly-rpoly}).
|
|
\begin{Lemma}\label{lem:tidb-reduce-poly}
|
|
For any \abbrCTIDB $\pdb$, $\raPlus$ query $\query$, and lineage polynomial
|
|
$\poly\inparen{\vct{X}}=\poly\pbox{\query,\tupset,\tup}\inparen{\vct{X}}$, it holds that $
|
|
\expct_{\vct{W} \sim \pdassign}\pbox{\refpoly{}\inparen{\vct{W}}} = \rpoly\inparen{\probAllTup}
|
|
$, where $\probAllTup = \inparen{\inparen{\prob_{\tup, j}}_{\tup\in\tupset, j\in\pbox{c}}}.$
|
|
\end{Lemma}
|
|
|
|
\subsection{Our Techniques}
|
|
\mypar{Lower Bound Proof Techniques}
|
|
Our main hardness result shows that computing~\Cref{prob:expect-mult} is $\sharpwonehard$ for $1$-\abbrTIDB. To prove this result we show that for the same $\query_1$ from the example above, for an arbitrary `product width' $k$, the query $\qhard^k$ is able to encode various hard graph-counting problems (assuming $\bigO{\numvar}$ tuples rather than the $\bigO{1}$ tuples in \Cref{fig:two-step}).
|
|
We do so by considering an arbitrary graph $G$ (analogous to relation $\boldsymbol{R}$ of $\query_1$) and analyzing how the coefficients in the (univariate) polynomial $\widetilde{\poly}\left(p,\dots,p\right)$ relate to counts of subgraphs in $G$ that are isomorphic to various graphs with $k$ edges. E.g., we exploit the fact that the coefficient corresponding to the power of $2k$ in $\poly$ of $\qhard^k$ is proportional to the number of $k$-matchings in $G$,
|
|
a known hard problem in parameterized/fine-grained complexity literature.
|
|
|
|
|
|
\mypar{Upper Bound Techniques}
|
|
Our negative results (\Cref{tab:lbs}) indicate that \abbrCTIDB{}s (even for $\bound=1$) can not achieve comparable performance to deterministic databases for exact results (under complexity assumptions). In fact, under plausible hardness conjectures, one cannot (drastically) improve upon the trivial algorithm to exactly compute the expected multiplicities for $1$-\abbrTIDB\xplural. A natural followup is whether we can do better if we are willing to settle for an approximation to the expected multiplities.
|
|
|
|
\input{two-step-model}
|
|
We adopt a two-step intensional model of query evaluation used in set-\abbrPDB\xplural, as illustrated in \Cref{fig:two-step}:
|
|
(i) \termStepOne (\abbrStepOne): Given input $\tupset$ and $\query$, output every tuple $\tup$ that possibly satisfies $\query$, annotated with its lineage polynomial ($\poly(\vct{X})=\apolyqdt\inparen{\vct{X}}$);
|
|
(ii) \termStepTwo (\abbrStepTwo): Given $\poly(\vct{X})$ for each tuple, compute $\expct_{\randWorld\sim\bpd}\pbox{\poly(\vct{\randWorld})}$.
|
|
Let $\timeOf{\abbrStepOne}(Q,\tupset,\circuit)$ denote the runtime of \abbrStepOne when it outputs $\circuit$ (which is a representation of $\poly$ as an arithmetic circuit --- more on this representation in~\Cref{sec:expression-trees}).
|
|
Denote by $\timeOf{\abbrStepTwo}(\circuit, \epsilon)$ (recall $\circuit$ is the output of \abbrStepOne) the runtime of \abbrStepTwo, which we can leverage~\Cref{def:reduced-poly} and~\Cref{lem:tidb-reduce-poly} to address the next formal objective:
|
|
|
|
\begin{Problem}[\abbrCTIDB linear time approximation]\label{prob:big-o-joint-steps}
|
|
Given \abbrCTIDB $\pdb$, $\raPlus$ query $\query$,
|
|
is there a $(1\pm\epsilon)$-approximation of $\expct_{\rvworld\sim\bpd}\pbox{\query\inparen{\rvworld}\inparen{\tup}}$ for all result tuples $\tup$ where
|
|
$\exists \circuit : \timeOf{\abbrStepOne}(Q,\tupset, \circuit) + \timeOf{\abbrStepTwo}(\circuit, \epsilon) \le O_\epsilon(\qruntime{\optquery{\query}, \tupset, \bound})$?
|
|
\end{Problem}
|
|
|
|
We show in \Cref{sec:circuit-depth} an $\bigO{\qruntime{\optquery{\query}, \tupset, \bound}}$ algorithm for constructing the lineage polynomial for all result tuples of an $\raPlus$ query $\query$ (or more more precisely, a single circuit $\circuit$ with one sink per tuple representing the tuple's lineage).
|
|
A key insight of this paper is that the representation of $\circuit$ matters.
|
|
For example, if we insist that $\circuit$ represent the lineage polynomial in \abbrSMB, the answer to the above question in general is no, since then we will need $\abs{\circuit}\ge \Omega\inparen{\inparen{\qruntime{\optquery{\query}, \tupset, \bound}}^k}$,
|
|
and hence, just $\timeOf{\abbrStepOne}(\query,\tupset,\circuit)$ will be too large.
|
|
|
|
However, systems can directly emit compact, factorized representations of $\poly(\vct{X})$ (e.g., as a consequence of the standard projection push-down optimization~\cite{DBLP:books/daglib/0020812}).
|
|
For example, in~\Cref{fig:two-step}, $B(Y+Z)$ is a factorized representation of the SMB-form $BY+BZ$.
|
|
Accordingly, this work uses (arithmetic) circuits\footnote{
|
|
An arithmetic circuit is a DAG with variable and/or numeric source nodes and internal, each nodes representing either an addition or multiplication operator.
|
|
}
|
|
as the representation system of $\poly(\vct{X})$.
|
|
|
|
Given that there exists a representation $\circuit^*$ such that $\timeOf{\abbrStepOne}(\query,\tupset,\circuit^*)\le \bigO{\qruntime{\optquery{\query}, \tupset, \bound}}$, we can now focus on the complexity of the \abbrStepTwo step.
|
|
We can represent the factorized lineage polynomial by its correspoding arithmetic circuit $\circuit$ (whose size we denote by $|\circuit|$).
|
|
As we also show in \Cref{sec:circuit-runtime}, this size is also bounded by $\qruntime{\optquery{\query}, \tupset, \bound}$ (i.e., $|\circuit^*| \le \bigO{\qruntime{\optquery{\query}, \tupset, \bound}}$).
|
|
Thus, the question of approximation
|
|
can be stated as the following stronger (since~\Cref{prob:big-o-joint-steps} has access to \emph{all} equivalent \circuit representing $\query\inparen{\vct{W}}\inparen{\tup}$), but sufficient condition:
|
|
\begin{Problem}\label{prob:intro-stmt}
|
|
Given one circuit $\circuit$ that encodes $\apolyqdt$ for all result tuples $\tup$ (one sink per $\tup$) for \abbrCTIDB $\pdb$ and $\raPlus$ query $\query$, does there exist an algorithm that computes a $(1\pm\epsilon)$-approximation of $\expct_{\rvworld\sim\bpd}\pbox{\query\inparen{\rvworld}\inparen{\tup}}$ (for all result tuples $\tup$) in $\bigO{|\circuit|}$ time?
|
|
\end{Problem}
|
|
|
|
For an upper bound on approximating the expected count, it is easy to check that if all the probabilties are constant then (with an additive adjustment) $\poly\left(\prob_1,\dots, \prob_n\right)$
|
|
|
|
(i.e. evaluating the original lineage polynomial
|
|
over the probability values) is a constant factor approximation
|
|
. This is illustrated in the following example using $\query_1^2$ from earlier. To aid in presentation we assume $\bound = 2$ for variable $X$ and $\bound = 1$ for all other variables. Let $\prob_A$ denote $\probOf\pbox{A = 1}$.
|
|
In computing $\rpoly$, we have some cancellations to deal with:
|
|
|
|
\begin{footnotesize}
|
|
\begin{align*}
|
|
\refpoly{1, }^2\inparen{\vct{X}} &= A^2\inparen{X_1^2 + 4X_1X_2 + 4X_2^2}B^2 + B^2Y^2E^2 + B^2Z^2C^2 + 2AX_1B^2YE \\
|
|
&\qquad+ 2AX_2B^2YE + 2AX_1B^2ZC + 2AX_2B^2ZC + 2B^2YEZC\\
|
|
\end{align*}
|
|
\end{footnotesize}
|
|
This then implies
|
|
\begin{footnotesize}
|
|
\begin{align*}
|
|
\rpoly^2\inparen{\vct{X}} &= AX_1B+4AX_2B+BYE+BZC+2AX_1BYE+2AX_2BYE+2AX_1BZC\\
|
|
&\qquad+2AX_2BZC+2BYEZC\\
|
|
\end{align*}
|
|
\end{footnotesize}
|
|
Substituting $\vct{\prob}$ for $\vct{X}$,
|
|
\begin{footnotesize}
|
|
\begin{align*}
|
|
\hspace*{-3mm}
|
|
\refpoly{1, }^2\inparen{\probAllTup} &= \prob_A^2\prob_{X_1}^2\prob_B^2 + 4\prob_A^2\prob_{X_1}\prob_{X_2}\prob_B^2 + 4\prob_A^2\prob_{X_2}^2\prob_B^2 + \prob_B^2\prob_Y^2\prob_E^2 + \prob_B^2\prob_Z^2\prob_C^2 + 2\prob_A\prob_{X_1}\prob_B^2\prob_Y\prob_E + 2\prob_A\prob_{X_2}\prob_B^2\prob_Y\prob_E\\
|
|
&\qquad+ 2\prob_A\prob_{X_1}\prob_B^2\prob_Z\prob_C + 2\prob_A\prob_{X_2}\prob_B^2\prob_Z\prob_C+ 2\prob_B^2\prob_Y\prob_E\prob_Z\prob_C\\
|
|
&\leq\prob_A\prob_{X_1}\prob_B + 4\prob_A^2\prob_{X_1}\prob_{X_2}\prob_B^2 + 4\prob_A\prob_{X_2}\prob_b + \prob_B\prob_Y\prob_E + \prob_B\prob_Z\prob_C + 2\prob_A\prob_{X_1}\prob_B\prob_Y\prob_E+ 2\prob_A\prob_{X_2}\prob_B\prob_Y\prob_E \\
|
|
&\qquad+ 2\prob_A\prob_{X_1}\prob_B\prob_Z\prob_C + 2\prob_A\prob_{X_2}\prob_B\prob_Z\prob_C + 2\prob_B\prob_Y\prob_E\prob_Z\prob_C
|
|
= \rpoly_1^2\inparen{\vct{p}} + 4\prob_A^2\prob_{X_1}\prob_{X_2}\prob_B^2.
|
|
\end{align*}
|
|
\end{footnotesize}
|
|
If we assume that all probability values are at least $p_0>0$, then given access to $\refpoly{1, }^2\inparen{\vct{\prob}} - 4\prob_A^2\prob_{X_1}\prob_{X_2}\prob_B^2$
|
|
we get that $\refpoly{1, }^2\inparen{\vct{\prob}} - 4\prob_A^2\prob_{X_1}\prob_{X_2}\prob_B^2$ is in the range $\left(\inparen{p_0}^3\cdot\inparen{\rpoly^2_1\vct{\prob}}, \rpoly_1^2\inparen{\vct{\prob}}\right]$.
|
|
%We can simulate sampling from $\refpoly{1, }^2\inparen{\vct{X}}$ by sampling monomials from $\refpoly{1, }^2$ while ignoring any samples $A^2X_1X_2B^2$.
|
|
Note however, that this is \emph{not a tight approximation}.
|
|
In~\cref{sec:algo} we demonstrate that a $(1\pm\epsilon)$ (multiplicative) approximation with competitive performance is achievable.
|
|
To get an $(1\pm \epsilon)$-multiplicative approximation and solve~\Cref{prob:intro-stmt}, using \circuit we uniformly sample monomials from the equivalent \abbrSMB representation of $\poly$ (without materializing the \abbrSMB representation) and `adjust' their contribution to $\widetilde{\poly}\left(\cdot\right)$.
|
|
|
|
|
|
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
|
|
|
\mypar{Applications}
|
|
Recent work in heuristic data cleaning~\cite{yang:2015:pvldb:lenses,DBLP:journals/vldb/SaRR0W0Z17,DBLP:journals/pvldb/RekatsinasCIR17,DBLP:journals/pvldb/BeskalesIG10,DBLP:journals/vldb/SaRR0W0Z17} emits a \abbrPDB when insufficient data exists to select the `correct' data repair.
|
|
Probabilistic data cleaning is a crucial innovation, as the alternative is to arbitrarily select one repair and `hope' that queries receive meaningful results.
|
|
Although \abbrPDB queries instead convey the trustworthiness of results~\cite{kumari:2016:qdb:communicating}, they are impractically slow~\cite{feng:2019:sigmod:uncertainty,feng:2021:sigmod:efficient}, even in approximation (see \Cref{sec:karp-luby}).
|
|
Bags, as we consider, are sufficient for production use, where bag-relational algebra is already the default for performance reasons.
|
|
Our results show that bag-\abbrPDB\xplural can be competitive, laying the groundwork for probabilistic functionality in production database engines.
|
|
|
|
|
|
|
|
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
|
\mypar{Paper Organization} We present relevant background and notation in \Cref{sec:background}. We then prove our main hardness results in \Cref{sec:hard} and present our approximation algorithm in \Cref{sec:algo}.
|
|
Finally, we discuss related work in \Cref{sec:related-work} and conclude in \Cref{sec:concl-future-work}. All proofs are in the appendix.
|
|
|
|
|
|
|
|
%%% Local Variables:
|
|
%%% mode: latex
|
|
%%% TeX-master: "main"
|
|
%%% End:
|