A query produces a result set consisting of rows, and each row consists of fields. There are several ways to receive this data.

The fields are untyped. That is to say, libpqxx has no opinion on what their respective types are. The database sends the data in a very flexible textual format. When you read a field, you specify what type you want it to be, and libpqxx converts the text format to that type for you.

If a value does not conform to the format for the type you specify, the conversion fails. For example, if you have strings that all happen to contain numbers, you can read them as int. But if any of the values is empty, or it's null (for a type that doesn't support null), or it's some string that does not look like an integer, or it's too large, you can't convert it to int.

So usually, reading result data from the database means not just retrieving the data; it also means converting it to some target type.

Querying rows of data

The simplest way to query rows of data is to call one of a transaction's "query" functions, passing as template arguments the types of columns you want to get back (e.g. int, std::string, double, and so on) and as a regular argument the query itself.

You can then iterate over the result to go over the rows of data:

for (auto [id, value] :
    tx.query<int, std::string>("SELECT id, name FROM item"))
{
    std::cout << id << '\t' << value << '\n';
}

The "query" functions execute your query, load the complete result data from the database, and iterate them. Along the way they convert each row to a tuple of C++ types that you indicated.

There are different query functions for querying any number of rows (query()); querying just one row of data as a std::tuple and throwing an error if it doesn't produce exactly one row (query1()); or querying either zero or one rows and getting a std::optional tuple of the values converted to the respective types you requested.

Streaming rows

There's another way to go through the rows coming out of a query. It's usually easier and faster if there are a lot of rows, but there are drawbacks.

One, you start getting rows before all the data has come in from the database. That speeds things up, but what happens if you lose your network connection while transferring the data? Your application may already have processed some of the data before finding out that the rest isn't coming. If that is a problem for your application, streaming may not be the right choice.

Two, streaming only works for some types of query. The stream() function wraps your query in a PostgreSQL COPY command, and COPY only supports a few commands: SELECT, VALUES, or an INSERT, UPDATE, or DELETE with a RETURNING clause. See the COPY documentation here: [ https://www.postgresql.org/docs/current/sql-copy.html ](https://www.postgresql.org/docs/current/sql-copy.html).

Three, when you convert a field to a "view" type (such as std::string_view or pqxx::bytes_view), the view points to underlying data which only stays valid until you iterate to the next row or exit the loop. So if you want to use that data for longer than a single iteration of the streaming loop, you'll have to store it somewhere yourself.

Now for the good news. Streaming does make it very easy to query data and loop over it, and often faster than with the "query" or "exec" functions:

for (auto [id, name, x, y] :
    tx.stream<int, std::string_view, float, float>(
        "SELECT id, name, x, y FROM point"))
  process(id + 1, "point-" + name, x * 10.0, y * 10.0);

The conversion to C++ types (here int, std::string_view, and two floats) is built into the function. You never even see row objects, field objects, iterators, or conversion methods. You just put in your query and you receive your data.

The 'result' class

Sometimes you want more from a query result than just rows of data. You may need to know how many rows your UPDATE statement has affected, or the names of the columns, and so on.

For that, use the transaction's exec() function. It returns a pqxx::result object. A result is really a smart pointer to a data structure managed by the C library, libpq. But seen from the outside, it acts like a random-access container of rows of data, which you can iterate like any other container, or index them like you would index an array. Each row in turn is a container of field values. And finally, each field holds a value.

Since result is at its core just a smart pointer, it is fairly cheap to copy: it doesn't copy the data, it just creates an extra reference to the same underlying data structure. When you destroy the last copy of the original result object you got back from the database, that will destroy the underlying data structure as a side effect.

Rows and fields

A field does not know what type of data it contains. It holds the data in string form. So, you specify the type when you read the value (using functions like as()), and the field will convert its string representation to that type.

By the way, there are no separate objects holding the rows or the fields. That data is all held inside the data structure managed by libpq. There are however two classes representing a row of data, and two classes representing a field:

_row_ref_: effectively a pointer to the result, plus a row number.
_row_: effectively a copy of the result, plus a row number.
_field_ref_: effectively a pointer to the result, with row & column numbers.
_field_: effectively a copy of the result, with row & column numbers.

When you index a result or dereference a result iterator, you get a row_ref. When you index a row_ref or row or dereference an iterator of either, you get a field_ref.

Usually you'll deal with row_ref and field_ref. These are effficient, cheap to copy, and straightforward. All you really need to do is ensure that the result object does not move in memory, or get destroyed.

In rare cases you may need to reference a row or field but without keeping the result, you can use a row or field object. These are less efficient, but they keep a result object so that the underlying data structure does not get destroyed so long as you stlil have a row or field referencing it.

Iterating rows and fields

For example, your code might just read it as raw text using c_str():

pqxx::result r = tx.exec("SELECT * FROM mytable");
for (auto const &row_ref: r)
{
   for (auto const &field_ref: row) std::cout << field.c_str() << '\t';
   std::cout << '\n';
}

Data types

Or you can ask for the field to convert itself to some other C++ type, using as<my_type>():

pqxx::result r = tx.exec("SELECT number, number * 2 FROM data");
for (auto const &row: r)
{
    for (auto const &field: row)
    {
        int n = field.as<int>();
        double f = field.as<double>();
        std::cout << n << '\t << f << '\n';
    }
}

Indexing

But results and rows also support other kinds of access. Array-style indexing, for instance, such as r[rownum]:

std::size_t const num_rows = std::size(r);
for (std::size_t rownum=0u; rownum < num_rows; ++rownum)
{
  pqxx::row_ref const row = r[rownum];
  std::size_t const num_cols = std::size(row);
  for (std::size_t colnum=0u; colnum < num_cols; ++colnum)
  {
    pqxx::field_ref const field = row[colnum];
    std::cout << field.c_str() << '\t';
  }
 
  std::cout << '\n';
}

Every row in the result has the same number of columns, so you don't need to look up the number of fields again for each one:

std::size_t const num_rows = std::size(r);
std::size_t const num_cols = r.columns();
for (std::size_t rownum=0u; rownum < num_rows; ++rownum)
{
  pqxx::row_ref const row = r[rownum];
  for (std::size_t colnum=0u; colnum < num_cols; ++colnum)
  {
    pqxx::field_ref const field = row[colnum];
    std::cout << field.c_str() << '\t';
  }
 
  std::cout << '\n';
}

You can even address a field by indexing the row using the field's name:

std::cout << row["salary"] << '\n';

But try not to do that if speed matters, because looking up the column by name takes time. At least you'd want to look up the column index before your loop and then use numerical indexes inside the loop.

For C++23 or better, there's also a two-dimensional array access operator, which is actually a little faster than the classic indexing:

for (std::size_t rownum=0u; rownum < num_rows; ++rownum)
{
    for (std::size_t colnum=0u; colnum < num_cols; ++colnum)
        std::cout result[rownum, colnum].c_str() << '\t';
    std::cout << '\n';
}

Going old-school

And of course you can use classic "begin/end" loops:

for (auto row = std::begin(r); row != std::end(r); row++)
{
  for (auto field = std::begin(row); field != std::end(row); field++)
    std::cout << field->c_str() << '\t';
  std::cout << '\n';
}

Fine print

Result sets are immutable, so all iterators on results and rows are actually const_iterators. There are also const_reverse_iterator types, which iterate backwards from rbegin() to rend() exclusive.

All these iterator types provide one extra bit of convenience that you won't normally find in C++ iterators: referential transparency. You don't need to dereference them to get to the row or field they refer to. That is, instead of row->end() you can also choose to say row.end(). Similarly, you may prefer field.c_str() over field->c_str().

This becomes really helpful with the array-indexing operator. With regular C++ iterators you would need ugly expressions like (*row)[0] or row->operator[](0). With the iterator types defined by the result and row classes you can simply say row[0].