-
Notifications
You must be signed in to change notification settings - Fork 979
BH Row Set Mechanism
(Submitted as a PR: /vector/src/main/java/org/apache/drill/exec/vector/accessor/RowSet.md
)
The row set mechanism is a handy, general purpose way to work with a batch of records. It came about from a desire to operator-level unit testing in which we create a batch of data, give it to the operator, and verify the resulting output. At that time, it was very difficult to create a batch of data: it required a large amount of complex code. (In production, this kind of code is generated, so we didn't have need to a simple interface.)
The term "row set" came about because of the overloaded nature of the term "batch" in Drill where it is both a set of rows and the operator that operates on those rows. The term "row set" is meant to invoke the idea of a collection of relational rows. (Also, the term was not previously used.)
The row set layer provides tools to work with all aspects of a batch life cycle:
- Define a schema (using the
SchemaBuilder
class) - Build a row set from a schema and a fluent description of the data (
RowSetBuilder
) - Attach to an existing vector container (
RowSet
and its implementations) - Print the row set (
print()
method on theRowSet
class) - Compare two row sets to verify test results (the awkwardly named
RowSetComparison
.)
The classes currently reside in src/main/test
. As least some of the classes may want to move to src/main/java
so they can be used to diagnose problems in a Drill server.
Drill has a whole menagerie of batch types:
-
RowSet
-
SingleRowSet
-
DirectRowSet
(single batch without a selection vector)-
ExtendableRowSet
(a writable single row set)
-
-
IndirectRowSet
(using an SV2)
-
-
HyperRowSet
(using an SV4)
-
Many operations are similar across row set types, while some differ. The row set class hierarchy captures those differences to provide a simple, consistent, easy-to-use API, primarily for tests.
The fluent schema builder can build all types of Drill schemas from the simplest (scalar types in a flat row) to the most complex (deeply nested complex types.)
(It is not entirely clear that the complex types earn their keep: they are very complex but only lightly used. They add tremendous complexity to the schema builder and all other parts of this project. We had to make them work, but one often wonders if they are really necessary.)
See ExampleTest
(or most of the tests in this project) for examples.
The row set writer is the little brother of the result set loader. The row set writer works on a single row set with a fixed schema. This limitation is fine in tests. Indeed, most of the time you will never actually see the row set writer: it is wrapped by the RowSetBuilder
which builds the vectors from schema, allocates the vectors, writes values to the vectors, and returns the fully populated row set.
For our purposes in this project, the row set writer serves two purposes:
- Allows us to create test batches easily.
- Provides a simple, pure test of the column writers without the added complexity of the result set loader.
Indeed, the row set writer uses exactly the same column writers as it's big brother, but without all the complexity of schema evolution, vector overflow, projection and the like. In short, if you change the column writers, test with the row set writer first before testing with the result set loader.
For obvious reasons, the row set writer works only with an extendable single row set (no indirection, no hyper vector.) There is a tool, however, to build up a hyper vector from a set of single row sets.
In some ways, the row set reader is simpler than the writer: it only need retrieve existing values. In other ways, the reader must handle more cases:
- Direct row set
- Indirect row set (using an SV2)
- Hyper row set (using an SV4)
The reader interface, however, is designed to hide all these details. The application simply creates a reader, steps through the rows, and reads out the values. The reader itself handles all necessary indirection.
The row set reader also works with a single batch. If we want to use this mechanism in operators, then we'll need a "result set reader" that handles multiple batches (along with any schema changes across batches.) That should be a very simple mechanism to build on top of the existing row set reader.
Most of the time you never use the row set reader directly. In a test, you want to know if a some "actual" batch (produced by an operator, say) matches an "expected" batch. The RowSetComparison
class does that work for you.
Both the reader and writer are designed for two distinct use cases:
- Type-aware code, such as that produced by code generation.
- Generic code, such as that that is convenient in tests.
Here, "generic" means (as a Java object) (not in the sense of Java Generics.)
The typed methods are much simpler than the "classic" complex writers which had a distinct "get" method for every data type. The readers and writers use a reduced set centered around the Java primitives. All types smaller than an int
are accessed as an int
(since Java uses that much space to pass values anyway.) Both FLOAT4
and FLOAT8
are returned as doubles. And so on.
(Date/time and interval classes are currently passed as Joda objects. For production code, this must be replaced with something more performant, such as returning fields in an array of int
s or some such.)
The generic methods are designed to enable simple generic code, such as that in RowSetComparison
: all values are treated as Java objects in both set
and get
methods. While we would never use these techniques in production, they same vast amount of complexity in tests.
Further, the accessor mechanism itself is design for generic (interpreted) access. The accessors are self describing so it is easy for test code to walk a column three for various tasks. (Something used again and again in test code.)