-
Notifications
You must be signed in to change notification settings - Fork 980
BH Metadata
(Submitted as a PR: /vector/src/main/java/org/apache/drill/exec/vector/accessor/Metadata.md)
Metadata is the glue that allows the many batch handling components to work together. There are several ways to categorize metadata:
- Descriptive: Describe existing data (that is, existing vectors or an existing table)
- Proscriptive: Describe the vectors we wish to build
At present, metadata groups into two distinct systems:
- Column metadata: used for proscriptive schemas and throughout the batch handling code. Provides many semantic services needed to describe, create and manage various kinds of metadata operations.
- Materialized field and batch schema: used to describe existing vectors and record batches (so-called vector containers.)
This section describes the purpose, services and role of each kind of metadata.
Drill already provides the MaterializedField class to describe a vector. This class has three components:
- Name
- Major type
- Minor (data) type
- Data mode (cardinality)
- Subtypes (for a union)
- Children (for a map, a repeated list or a list)
The materialized field is straightforward for simple types: INT, VARCHAR, INTERVAL and so on. The original design appears to be that, because these types are simple, a MaterializedField is immutable. Quite simple.
Drill also supports complex types: those that are made up of one column that contains other columns. In Drill, the complex types are; Map, Repeated Map, (non-repeated) List, Repeated List and Union. In this case, MaterializedField must describe not only the complex vector itself, but also the children. (A map must describe its members, for example.)
This requirement puts great strain on the concept that a MaterializedField is immutable. A map, say, allows adding new members at any time. How do we add new children to the map's MaterializedField if it is immutable? The solution appears to be that the name and type are immutable, but the list of children is not. It is a compromise, but it works.
Unions (though they are barely supported) introduce another issue. A union contains a series of "subtypes" (vectors) that represent the data of the union. Each row contains a value in one of the subtypes: in, say, INT or VARCHAR. Unions allow the addition of subtypes at any time. Since the MajorType and MaterializedField are immutable, now do we handle this? The answer has been to create a new MaterializedField for the union vector.
Now things get interesting. A map can contain a union. The map's MaterializedField holds a list of children, including the MaterializedField for the union. If we replace the union's MaterializedField the map will then hold onto a stale version and the schema becomes inconsistent.
The code for this project tries to find and fix these issues where they are found. (But, we really should test every pair of complex types to ensure that all cases are handled.) In general, the solution is to discard the assumption that the MaterializedField is immutable, replacing with a finer-grain set of invariants (such as that neither the name nor the data type/mode can change.)
A nullable INT vector is, conceptually just a nullable INT type an a name. But, the nullable vector is represented as a container with children for the "bits" and "data" vectors. The actual MaterializedField contains these hidden vectors as children. But, from a conceptual point of view, the children are implementation details. One key place where this issue arrises is in comparison. Is a nullable INT (without children) equal to a nullable INT with children? For some purposes, yes. For others, no.
The recently-added isEquivalent() method on MaterializedField resolves this ambiguity by considering children only for types in which children are user-visible (maps, unions, etc.)
The (obscure) UnionVector is a single vector that can hold multiple types. A row is one type or another. The structure here is a bit unclear. What is clear is that, for a union vector, the child types are stored as subtypes in the MajorType. The children are also stored, as materialized fields, but in an awkward structure. Like a nullable vector, union vector has internal structure stored as children in the the materialized field. The actual type vectors are stored two levels down, as children of an internal map that is the child of the union.
This structure works, but is very awkward for semantic processing because of the amount of implementation detail that "shows through" into the metadata.
Drill has two kinds of lists. The "non-repeated" (and basically unsupported) ListVector and the RepeatedListVector. Despite the similar names, the are very separate concepts.
The list vector is quite odd. It has a "bits" vector that allows arrays to be marked as null (as is needed for JSON.) It also can act as either a "nullable repeated nullable X" or as a "repeated union". The materialized field thus has a very complex structure.
When acting as "nullable repeated nullable X", the materialized field for the "nullable X" appears as a child.
But, when the list is "promoted to union", then the child structure changes. In this case, the union is the child of the list, and the "nullable X" is pushed two levels down as described for unions above.
Again, this works, but becomes quite vexing for semantic processing.
The MaterializedField has a number of other limitations that make it awkward to use for semantic processing:
- No convenient access to the most important properties: data type and cardinality. Each access requires an awkward chain of calls.
- Inconvenient equality checks: equality is based on the equality of Protobufs, which have very odd semantics (such as whether a field is set or not.) Also, there is ambiguity with the hidden children.
- Very awkward access to children (only an iterator is provided, not indexed access.)
To be fair, MaterializedField was never meant for the kind of semantic processing we need. All of this is not meant as a critique of the code (the code is fine), but rather as a motivation for the enhanced metadata mechanism described below.
Drill represents a record batch (the data, not the operator) as a VectorContainer. The container gathers up the schemas of its vectors into a BatchSchema. However, the BatchSchema is a bit awkward, it is missing methods needed for semantic processing.
For these reasons (and more), this project decided to create a new metadata representation: one that is better suited to semantic processing. One of the key design goals of this approach was to realize a very helpful generalization: rows and maps are both tuples and should have the same metadata description.
The metadata layer is defined in the vector module alongside the MaterializedField class, vectors and column accessors (described later). However, they are implemented in the java-exec package, which is where the BatchSchema is defined.
The two key interfaces are:
-
TupleMetadata- generalized representation of a row or map. -
ColumnMetadata- generalized representation of a named column.
Two additional interfaces exist for Drill's obscure data types:
-
VariantMetadata- represents a Union or a (non-repeated) List
The column/tuple metadata combine to form a simple tree structure, rooted in a tuple, that can describe (possibly repeated) maps to any level of depth. Include unions, and the tree can represent any combination of nesting of maps in unions, unions in maps, and so on.
Traditional databases employ the "tuple" concept. In practice, a tuple is a collection of values identified by both name an position. Consider JDBC as a typical example, we can access a column either by name ("myCol", say) or by position (4, say). Today Drill is a bit conflicted on indexed access. VectorContainer provides indexed access, as do maps. Unions do not, nor do MaterializedField child lists. Drill also provides name access, but it is implemented as a linear search in some cases.
Since both name an indexed lookups are vital (name lookup for semantic processing, index for performant runtime lookups), the tuple metadata (as well as the column accessors) provide both indexed and name access. Name access is via a hash map for performance.
In addition to the row/map unification, the column metadata helps to smooth over a number of other awkward bits in the MaterializedField representation:
- The tuple schema reifies the concept of a map schema in a form much more convenient than the
MaterializedFieldchild iterator. The tuple schema provides services to easily build, traverse and query a row or map schema. - Many query methods for common conditions.
- Divide columns up into four broad groups: Primitive, Tuple, Variant and Multi-Array. (The last represents a repeated list.)
The key job of the metadata interfaces are to provide semantic services such as names, types, cardinality, default data width, default array sizes, and more. In the implementation package, the RowSet version of the SchemaBuilder provides a fluent way to define a schema for tests. Once you understand the purpose of these classes, the JavaDoc in the class should provide the necessary details.
One non-obvious service is to report, in the context of a scan operator, if a column is projected. In this way, a reader can declare it's view of the schema. (A CSV reader, say, can declare that it has 10 columns.) After schema negotiation, the scan operator will determine that only three of the ten columns are projected. The column metadata for the table (the CSV file) will identify which are projected into the scan operator's output schema. (More on this concept later.)
When allocating a new vector, we often need to know up to three pieces of information:
- The number of rows
- The width (for VarChar fields)
- The inner size (for arrays)
Given this information, Drill can very accurately allocate a vector of the required length. Doing so is important because it eliminates the doubling and copying required with poor guesses. Most Drill code provides only the number of rows, and guesses for the others. (This is where the infamous VarChar size of 50 comes in, along with a guess of 5 for array sizes.) The sort operator took an initial stab at an advanced batch allocator that uses the additional values (which it obtained by introspecting an incoming batch using the BatchSizer.)
This mechanism goes a step further and adds the additional values to the metadata itself. Over time, readers can determine the value through inspection, and can (we hope) pass the values along with each batch so that other operators don't need to repeated run the BatchSizer over and over. But, that is a long-term goal.
Tools exist to examine a batch (using the MaterializedFields) and create a the batch metadata. This is done, for example, in the RowSet classes to provide a simple schema description which is then used to generate readers or writers for that batch.
Other tools (also in the RowSet tools) can build a batch of vectors from a schema. Thus, the column metadata provide complete "round trip" engineering. (However, at present, the additional fields (width and array size) are lost in the process since MaterializedField does not support them.)
The scan operator implements projection push-down in which the query (via the project list) "negotiates" with the data source (via the reader) for the of columns to provide. The scanner projects some, creates nulls for others, and implicit columns for others. In the new framework, all of this work uses the column/tuple metadata to represent a schema.
Obviously it is less than ideal to have two distinct metadata systems. A good long-term goal would be to merge the two systems. In particular, MaterializedField would be replaced by the column metadata, and the BatchSchema would be replaced by the tuple schema.
The rub, of course, is that such a change breaks the client/server API. Changing value vector metadata would require changes to the JDBC and ODBC clients to handle the new format. Without API versioning, this is a difficult change. So, for now, we use two systems as that approach, while awkward, does have the advantage of not breaking backward compatibility.