Reading and Writing CSV files#

Arrow provides a fast CSV reader allowing ingestion of external data to create Arrow Tables or a stream of Arrow RecordBatches.

Reading CSV files#

Data in a CSV file can either be read in as a single Arrow Table using TableReader or streamed as RecordBatches using StreamingReader. See Tradeoffs for a discussion of the tradeoffs between the two methods.

Both these readers require an arrow::io::InputStream instance representing the input file. Their behavior can be customized using a combination of ReadOptions, ParseOptions, and ConvertOptions.

TableReader#

#include "arrow/csv/api.h"

{
   // ...
   arrow::io::IOContext io_context = arrow::io::default_io_context();
   std::shared_ptr<arrow::io::InputStream> input = ...;

   auto read_options = arrow::csv::ReadOptions::Defaults();
   auto parse_options = arrow::csv::ParseOptions::Defaults();
   auto convert_options = arrow::csv::ConvertOptions::Defaults();

   // Instantiate TableReader from input stream and options
   auto maybe_reader =
     arrow::csv::TableReader::Make(io_context,
                                   input,
                                   read_options,
                                   parse_options,
                                   convert_options);
   if (!maybe_reader.ok()) {
     // Handle TableReader instantiation error...
   }
   std::shared_ptr<arrow::csv::TableReader> reader = *maybe_reader;

   // Read table from CSV file
   auto maybe_table = reader->Read();
   if (!maybe_table.ok()) {
     // Handle CSV read error
     // (for example a CSV syntax error or failed type conversion)
   }
   std::shared_ptr<arrow::Table> table = *maybe_table;
}

StreamingReader#

#include "arrow/csv/api.h"

{
   // ...
   arrow::io::IOContext io_context = arrow::io::default_io_context();
   std::shared_ptr<arrow::io::InputStream> input = ...;

   auto read_options = arrow::csv::ReadOptions::Defaults();
   auto parse_options = arrow::csv::ParseOptions::Defaults();
   auto convert_options = arrow::csv::ConvertOptions::Defaults();

   // Instantiate StreamingReader from input stream and options
   auto maybe_reader =
     arrow::csv::StreamingReader::Make(io_context,
                                       input,
                                       read_options,
                                       parse_options,
                                       convert_options);
   if (!maybe_reader.ok()) {
     // Handle StreamingReader instantiation error...
   }
   std::shared_ptr<arrow::csv::StreamingReader> reader = *maybe_reader;

   // Set aside a RecordBatch pointer for re-use while streaming
   std::shared_ptr<RecordBatch> batch;

   while (true) {
       // Attempt to read the first RecordBatch
       arrow::Status status = reader->ReadNext(&batch);

       if (!status.ok()) {
         // Handle read error
       }

       if (batch == NULL) {
         // Handle end of file
         break;
       }

       // Do something with the batch
   }
}

Tradeoffs#

The choice between using TableReader or StreamingReader will ultimately depend on the use case but there are a few tradeoffs to be aware of:

Memory usage: TableReader loads all of the data into memory at once and, depending on the amount of data, may require considerably more memory than StreamingReader which only loads one RecordBatch at a time. This is likely to be the most significant tradeoff for users.
Speed: When reading the entire contents of a CSV, TableReader will tend to be faster than StreamingReader because it makes better use of available cores. See Performance for more details.
Flexibility: StreamingReader might be considered less flexible than TableReader because it performs type inference only on the first block that’s read in, after which point the types are frozen and any data in subsequent blocks that cannot be converted to those types will cause an error. Note that this can be remedied either by setting ReadOptions::block_size to a large enough value or by using ConvertOptions::column_types to set the desired data types explicitly.

Writing CSV files#

A CSV file is written to a OutputStream.

#include <arrow/csv/api.h>
{
    // Oneshot write
    // ...
    std::shared_ptr<arrow::io::OutputStream> output = ...;
    auto write_options = arrow::csv::WriteOptions::Defaults();
    if (WriteCSV(table, write_options, output.get()).ok()) {
        // Handle writer error...
    }
}
{
    // Write incrementally
    // ...
    std::shared_ptr<arrow::io::OutputStream> output = ...;
    auto write_options = arrow::csv::WriteOptions::Defaults();
    auto maybe_writer = arrow::csv::MakeCSVWriter(output, schema, write_options);
    if (!maybe_writer.ok()) {
        // Handle writer instantiation error...
    }
    std::shared_ptr<arrow::ipc::RecordBatchWriter> writer = *maybe_writer;

    // Write batches...
    if (!writer->WriteRecordBatch(*batch).ok()) {
        // Handle write error...
    }

    if (!writer->Close().ok()) {
        // Handle close error...
    }
    if (!output->Close().ok()) {
        // Handle file close error...
    }
}

Note

The writer does not yet support all Arrow types.

Column names#

There are three possible ways to infer column names from the CSV file:

By default, the column names are read from the first row in the CSV file
If ReadOptions::column_names is set, it forces the column names in the table to these values (the first row in the CSV file is read as data)
If ReadOptions::autogenerate_column_names is true, column names will be autogenerated with the pattern “f0”, “f1”… (the first row in the CSV file is read as data)

Column selection#

By default, Arrow reads all columns in the CSV file. You can narrow the selection of columns with the ConvertOptions::include_columns option. If some columns in ConvertOptions::include_columns are missing from the CSV file, an error will be emitted unless ConvertOptions::include_missing_columns is true, in which case the missing columns are assumed to contain all-null values.

Interaction with column names#

If both ReadOptions::column_names and ConvertOptions::include_columns are specified, the ReadOptions::column_names are assumed to map to CSV columns, and ConvertOptions::include_columns is a subset of those column names that will part of the Arrow Table.

Data types#

By default, the CSV reader infers the most appropriate data type for each column. Type inference considers the following data types, in order:

Null
Int64
Boolean
Date32
Time32 (with seconds unit)
Timestamp (with seconds unit)
Timestamp (with nanoseconds unit)
Float64
Dictionary<String> (if ConvertOptions::auto_dict_encode is true)
Dictionary<Binary> (if ConvertOptions::auto_dict_encode is true)
String
Binary

It is possible to override type inference for select columns by setting the ConvertOptions::column_types option. Explicit data types can be chosen from the following list:

Null
All Integer types
Float32 and Float64
Decimal128
Boolean
Date32 and Date64
Time32 and Time64
Timestamp
Binary and Large Binary
String and Large String (with optional UTF8 input validation)
Fixed-Size Binary
Dictionary with index type Int32 and value type one of the following: Binary, String, LargeBinary, LargeString, Int32, UInt32, Int64, UInt64, Float32, Float64, Decimal128

Other data types do not support conversion from CSV values and will error out.

Dictionary inference#

If type inference is enabled and ConvertOptions::auto_dict_encode is true, the CSV reader first tries to convert string-like columns to a dictionary-encoded string-like array. It switches to a plain string-like array when the threshold in ConvertOptions::auto_dict_max_cardinality is reached.

Timestamp inference/parsing#

If type inference is enabled, the CSV reader first tries to interpret string-like columns as timestamps. If all rows have some zone offset (e.g. Z or +0100), even if the offsets are inconsistent, then the inferred type will be UTC timestamp. If no rows have a zone offset, then the inferred type will be timestamp without timezone. A mix of rows with/without offsets will result in a string column.

If the type is explicitly specified as a timestamp with/without timezone, then the reader will error on values without/with zone offsets in that column. Note that this means it isn’t currently possible to have the reader parse a column of timestamps without zone offsets as local times in a particular timezone; instead, parse the column as timestamp without timezone, then convert the values afterwards using the assume_timezone compute function.

Specified Type	Input CSV	Result Type
(inferred)	`2021-01-01T00:00:00`	timestamp[s]
	`2021-01-01T00:00:00Z`	timestamp[s, UTC]
	`2021-01-01T00:00:00+0100`	timestamp[s, UTC]
	2021-01-01T00:00:00 2021-01-01T00:00:00Z	string
timestamp[s]	`2021-01-01T00:00:00`	timestamp[s]
	`2021-01-01T00:00:00Z`	(error)
	`2021-01-01T00:00:00+0100`
	2021-01-01T00:00:00 2021-01-01T00:00:00Z
timestamp[s, UTC]	`2021-01-01T00:00:00`	(error)
	`2021-01-01T00:00:00Z`	timestamp[s, UTC]
	`2021-01-01T00:00:00+0100`	timestamp[s, UTC]
	2021-01-01T00:00:00 2021-01-01T00:00:00Z	(error)
timestamp[s, America/New_York]	`2021-01-01T00:00:00`	(error)
	`2021-01-01T00:00:00Z`	timestamp[s, America/New_York]
	`2021-01-01T00:00:00+0100`	timestamp[s, America/New_York]
	2021-01-01T00:00:00 2021-01-01T00:00:00Z	(error)

Nulls#

Null values are recognized from the spellings stored in ConvertOptions::null_values. The ConvertOptions::Defaults() factory method will initialize a number of conventional null spellings such as N/A.

Character encoding#

CSV files are expected to be encoded in UTF8. However, non-UTF8 data is accepted for Binary columns.

Write Options#

The format of written CSV files can be customized via WriteOptions. Currently few options are available; more will be added in future releases.

Performance#

By default, TableReader will parallelize reads in order to exploit all CPU cores on your machine. You can change this setting in ReadOptions::use_threads. A reasonable expectation is at least 100 MB/s per core on a performant desktop or laptop computer (measured in source CSV bytes, not target Arrow data bytes).