[
https://issues.apache.org/jira/browse/DRILL-5376?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15940797#comment-15940797
]
Paul Rogers commented on DRILL-5376:
------------------------------------
Another asymmetry between rows and maps occurs in the "hyper vector." The
simplest hyper vector is simply two value vectors "stacked" one atop another.
An SV4 indexes into the two separate vectors.
{code}
| 10 |
| 20 |
| 30 |
--
| 40 |
| 50 |
{code}
A hyper vector of row batches is a set of stacked vectors. Very simple to index
into. An SV4 is essentially a pair (batch, offset), where batch gives the index
of the vector in the stack, the offset gives the offset of the data within the
vector.
But, a map hyper vector has a completely different structure. It is a stack of
maps, each map with a list of its vectors.
{code}
| map1 --> ([10, 20, 30], ["a", "b", "c"]) |
- - -
| map2 --> ([1, 2, 3], ["fred", "betty", "wilma"]) |
{code}
Indexing here is much harder. Assume the schema is (map( a, b)). That is a map
field. In the map are fields a and b. To get to the third a value in the second
batch (2, 3), one uses the batch to find map2. One does not index into map2 as
above, instead, one then looks up field a to find the corresponding value
vector. Then one indexes into vector a.
The complexity mounts with nested maps.
Instead, a very easy alternative is simply to stack the map members:
{code}
| 20 | "a" |
| 30 | "b" |
| 40 | "c" |
- - - -
| 1 | "fred" |
| 2 | "betty" |
| 3 | "wilma" |
{code}
Now, to get to (3,2) for a, one simply finds the "map.a" vector, and indexes
into it the same as for top-level scalar vectors. No need for two distinct code
paths.
> Rationalize Drill's row metadata for simpler code, better performance
> ---------------------------------------------------------------------
>
> Key: DRILL-5376
> URL: https://issues.apache.org/jira/browse/DRILL-5376
> Project: Apache Drill
> Issue Type: Improvement
> Affects Versions: 1.10.0
> Reporter: Paul Rogers
>
> Drill is a columnar system, but data is ultimately represented as rows (AKA
> records or tuples.) The way that Drill represents rows leads to excessive
> code complexity and runtime cost.
> Data in Drill is stored in vectors: one (or more) per column. Vectors do not
> stand alone, however, they are "bundled" into various forms of grouping: the
> {{VectorContainer}}, {{RecordBatch}}, {{VectorAccessible}},
> {{VectorAccessibleSerializable}}, and more. Each has slightly different
> semantics, requiring large amounts of code to bridge between the
> representations.
> Consider only a simple row: one with only scalar columns. In classic
> relational theory, such a row is a tuple:
> {code}
> R = (a, b, c, d, ...)
> {code}
> A tuple is defined as an ordered list of column values. Unlike a list or
> array, the column values also have names and may have varying data types.
> In SQL, columns are referenced by either position or name. In most execution
> engines, columns are referenced by position (since positions, in most
> systems, cannot change.) A 1:1 mapping is provided between names and
> positions. (See the JDBC {{RecordSet}} interface.)
> This allows code to be very fast: code references columns by index, not by
> name, avoiding name lookups for each column reference.
> Drill provides a murky, hybrid approach. Some structures ({{BatchSchema}},
> for example) appear to provide a fixed column ordering, allowing indexed
> column access. But, other abstractions provide only an iterator. Others (such
> as {{VectorContainer}}) provides name-based access or, by clever programming,
> indexed access.
> As a result, it is never clear exactly how to quickly access a column: by
> name, by name to multi-part index to vector?
> Of course, Drill also supports maps, which add to the complexity. First, we
> must understand that a "map" in Drill is not a "map" in the classic sense: it
> is not a collection of (name, value) pairs in the JSON sense: a collection in
> which each instance may have a different set of pairs.
> Instead, in Drill, a "map" is really a nested tuple: a map has the same
> structure as a Drill record: a collection of names and values in which all
> rows have the same structure. (This is so because maps are really a
> collection of value vectors, and the vectors cut across all rows.)
> Drill, however, does not reflect this symmetry: that a row and a map are both
> tuples. There are no common abstractions for the two. Instead, maps are
> represented as a {{MapVector}} that contains a (name, vector) map for its
> children.
> Because of this name-based mapping, high-speed indexed access to vectors is
> not provided "out of the box." Certainly each consumer of a map can build its
> own indexing mechanism. But, this leads to code complexity and redundancy.
> This ticket asks to rationalize Drill's row, map and schema abstractions
> around the tuple concept. A schema is a description of a tuple and should (as
> in JDBC) provide both name and index based access. That is, provide methods
> of the form:
> {code}
> MaterializedField getField(int index);
> MaterializedField getField(String name);
> ...
> ValueVector getVector(int index);
> ValueVector getVector(String name);
> {code}
> Provide a common abstraction for rows and maps, recognizing their structural
> similarity.
> There is an obvious issue with indexing columns in a row when the row
> contains maps. Should indexing be multi-part (index into row, then into map)
> as today? A better alternative is to provide a flattened interface:
> {code}
> 0: a, 1: b.x, 2: b.y, 3: c, ...
> {code}
> Use this change to simplify client code, over time, to use a simple
> indexed-based column access.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
