HDK is a low-level execution library for data analytics processing. HDK is used as a fast execution backend in Modin. The HDK library provides a set of components for federating analytic queries to an execution backend based on OmniSciDB.
OmniSciDB is an open-source SQL-based relational database designed for the massive parallelism of modern CPU and GPU hardware. Its execution engine is built on LLVM JIT compiler.
HDK can be embedded into an application as a python module -
pyhdk. This module
provides Python APIs to the HDK library. A specialized in-memory storage layer
provides an efficient way to import data in Arrow table format.
HdkOnNative execution uses HDK for both as a storage format and for actual data transformation.
Relational engine limitations#
Using a relational database engine implies a set of restrictions on operations we can execute on a dataframe.
We cannot handle frames that use data types not supported by OmniSciDB. Currently, we allow only integer, float, string, and categorical data types.
Column data should be homogeneous.
Can only support operations that map to relational algebra. This means most operations are supported over a single axis (axis=0) only. Non-relational operations like transposition and pivot are not supported.
When the unsupported data type is detected or unsupported operations is requested it falls back to the original pandas framework.
In Modin, partitioning is used to achieve high parallelism. In the case of
HDK-based execution, parallelism is provided by HDK execution
engine itself and we don’t need to manage multiple partitions.
always has a single partition.
A partition holds data in either
pandas.DataFrame is preferred only when we detect unsupported
data type and therefore have to use
pandas framework for processing.
In other cases
pyarrow.Table format is preferred. Arrow tables can be
zero-copy imported into HDK. A query execution result is also
returned as an Arrow table.
When users import data in Modin DataFrame (from a file or from some python
object like array or dictionary) they invoke one of the
functions (to read data from a file) or use
(to create a DataFrame from an iterable object). Both of the paths lead to the
that defines a factory that handles the import query. For HdkOnNative
execution, the factory is accordingly
The factory dispatches the import query: if the data needs to be read from a file
- the query is routed to the
class, that uses Arrow Framework to read the file into a PyArrow Table, the resulted
table is passed to the
If the factory deals with importing a Python’s iterable object, the query goes straight
The Query Compiler sanitizes an input object and passes it to one of the
factory methods (
.from_*). The Dataframe’s build method stores the passed object into a new Dataframe’s partition
and returns the resulted Dataframe, which is then wrapped into a Query Compiler, which is
wrapped into a high-level Modin DataFrame, which is returned to the user.
Note that during this ingress flow, no data is actually imported to HDK. The need for
importing to HDK is decided later at the execution stage by the Modin Core Dataframe layer.
If the query requires for the data to be placed in HDK, the import is triggered.
passes partition to import to the
that extracts a partition’s underlying object and sends a request to import it to HDK.
The response for the request is a unique identifier for the just imported table
at HDK, this identifier is placed in the partition. After that, the partition has
a reference to the concrete table in HDK to query, and the data is considered to be
When a user calls any
DataFrame API, a query
starts forming at the API layer to be executed at the Execution layer. The API
layer is responsible for processing the query appropriately, for example, determining
whether the final result should be a
Series object, and
sanitizing the input to the
e.g. validating a parameter from the query and defining specific intermediate values
to provide more context to the query compiler.
is responsible for reducing the query to the pre-defined Dataframe algebra operators
and triggering execution on the
receives a query, it determines whether the operation requires data materialization
or whether it can be performed lazily. The operation is then either appended to a
lazy computation tree or executed immediately.
HDK has a powerful query optimizer and an execution engine that combines multiple operations into a single execution module. E.g. join, filter and aggregation can be executed in a single data scan.
To utilize this feature and reduce data transformation and transfer overheads, all of the operations that don’t require data materialization are performed lazily.
Lazy operations on a frame build a tree which is later translated into a query executed by HDK. Each of the tree nodes has its input node(s) - a frame argument(s) of the operation. When a new node is appended to the tree, it becomes its root. The leaves of the tree are always a special node type, whose input is an actual materialized frame to execute operations from the tree on.
There are two types of trees. The first one describes operations on frames that
map to relational operations like projection, union, etc. Nodes in this tree are
class. Leaf nodes are instances of the
class. The second type of tree is used to describe operations on columns, including
arithmetic operations, type casts, datetime operations, etc. Nodes of this tree are derived from
class. Leaf nodes are instances of the
Visit the corresponding sections to go through all of the types of nodes:
Execution of a computation tree#
Frames are materialized (executed) when their data is accessed. E.g. it happens when we try to access the frame’s index or shape. There are two ways to execute required operations: through Arrow or through HDK.
For simple operations which don’t include actual computations, execution can use Arrow API. We can use it to rename columns, drop columns and concatenate frames. Arrow execution is preferable since it doesn’t require actual data import/export from/to HDK.
To execute a query in the HDK engine we need to import data first. We should find all leaves of an operation tree and import their Arrow tables. Partitions with imported tables hold corresponding table names used to refer to them in queries.
HDK executes queries expressed in HDK-specific intermediate representation (IR) format.
It also provides components to translate SQL queries to relational algebra JSON format
which can be later optimized and translated to HDK IR. Modin generates queries in relational
algebra JSON format. These queries are optionally optimized with Apache Calcite
based optimizer provided by HDK (
Calcite) and then executed.
Operations used by Calcite in its intermediate representation are implemented
in classes derived from
CalciteBuilder is used to
It also translates
trees by replacing
nodes with either
depending on context.
is used to serialize the resulting sequence into
JSON format. This JSON becomes a query by simply adding ‘execute relalg’
or ‘execute calcite’ prefix (the latter is used if we want to use Calcite
for additional query optimization).
The building of Calcite query (starting from the conversion to the Calcite Algebra and up to
the forming JSON query) is orchestrated by
An execution result is a new Arrow table which is used to form a new
partition. This partition is assigned to the executed frame. The frame’s
operation tree is replaced with
Rowid column and sub-queries#
A special case of an index is the default index - 0-based numeric sequence.
In our representation, such an index is represented by the absence of index columns.
If we need to access the index value we can use the virtual
rowid column provided
by HDK. Unfortunately, this special column is available for physical
tables only. That means we cannot access it for a node that is not a tree leaf.
That makes us execute trees with such nodes in several steps. First, we
materialize all frames that require
rowid column and only after that we can
materialize the root of the tree.
HdkOnNative Dataframe Implementation#
To support lazy execution Modin uses two types of trees. Operations on frames are described
DFAlgNode based trees. Scalar computations are described by
BaseExpr based tree.
Interactions with HDK engine are done using
HdkWorker class. Queries use serialized
Calcite relational algebra format. Calcite algebra nodes are based on
Translation is done by
CalciteBuilder class. Serialization is performed by
Column name mangling#
pandas.DataFrame columns might have names not allowed in SQL (e. g.
an empty string). To handle this we simply add ‘F_’ prefix to
column names. Index labels are more tricky because they might be non-unique.
Indexes are represented as regular columns, and we have to perform a special
mangling to get valid and unique column names. Demangling is done when we
transform our frame (i.e. its Arrow table) into