Erlang PostgreSQL Database Client

Asynchronous fork of wg/epgsql originally here: mabrek/epgsql and subsequently forked in order to provide a common fork for community development.

pgapp

If you want to get up to speed quickly with code that lets you run Postgres queries, you might consider trying epgsql/pgapp, which adds the following, on top of the epgsql driver:

A 'resource pool' (currently poolboy), which lets you decide how many Postgres workers you want to utilize.
Resilience against the database going down or other problems. The pgapp code will keep trying to reconnect to the database, but will not propagate the crash up the supervisor tree, so that, for instance, your web site will stay up even if the database is down for some reason. Erlang's "let it crash" is a good idea, but external resources going away might not be a good reason to crash your entire system.

Motivation

When you need to execute several queries, it involves a number network round-trips between the application and the database. The PostgreSQL frontend/backend protocol supports request pipelining. This means that you don't need to wait for the previous command to finish before sending the next command. This version of the driver makes full use of the protocol feature that allows faster execution.

Difference highlights

3 API sets:
- epgsql maintains backwards compatibility with the original driver API
- epgsqla delivers complete results as regular erlang messages
- epgsqli delivers results as messages incrementally (row by row) All API interfaces can be used with the same connection: eg, connection opened with epgsql can be queried with epgsql / epgsqla / epgsqli in any combinations.
internal queue of client requests, so you don't need to wait for the response to send the next request (pipelining)
single process to hold driver state and receive socket data
execution of several parsed statements as a batch
binding timestamps in erlang:now() format

see CHANGES for full list.

Differences between current epgsql and mabrek's original async fork:

Unnamed statements are used unless specified otherwise. This may cause problems for people attempting to use the same connection concurrently, which will no longer work.

Known problems

SSL performance can degrade if the driver process has a large inbox (thousands of messages).

Usage

Connect

connect(Opts) -> {ok, Connection :: epgsql:connection()} | {error, Reason :: epgsql:connect_error()}
    when
  Opts ::
    #{host :=     inet:ip_address() | inet:hostname(),
      username := iodata(),
      password => iodata() | fun( () -> iodata() ),
      database => iodata(),
      port =>     inet:port_number(),
      ssl =>      boolean() | required,
      ssl_opts => [ssl:tls_client_option()], % @see OTP ssl documentation
      socket_active => true | integer(), % @see "Active socket" section below
      tcp_opts => [gen_tcp:option()],    % @see OTP gen_tcp module documentation
      timeout =>  timeout(),             % socket connect timeout, default: 5000 ms
      async =>    pid() | atom(),        % process to receive LISTEN/NOTIFY msgs
      codecs =>   [{epgsql_codec:codec_mod(), any()}]}
      nulls =>    [any(), ...],          % NULL terms
      replication => Replication :: string()} % Pass "database" to connect in replication mode
    | list().

connect(Host, Username, Password, Opts) -> {ok, C} | {error, Reason}.

example:

{ok, C} = epgsql:connect(#{
    host => "localhost",
    username => "username",
    password => "psss",
    database => "test_db",
    timeout => 4000
}),
...
ok = epgsql:close(C).

Only host and username are mandatory, but most likely you would need database and password.

password - DB user password. It might be provided as string / binary or as a fun that returns string / binary. Internally, plain password is wrapped to anonymous fun before it is sent to connection process, so, if connect command crashes, plain password will not appear in crash logs.
timeout parameter will trigger an {error, timeout} result when the socket fails to connect within provided milliseconds.
ssl if set to true, perform an attempt to connect in ssl mode, but continue unencrypted if encryption isn't supported by server. if set to required connection will fail if encryption is not available.
ssl_opts will be passed as is to ssl:connect/3.
tcp_opts will be passed as is to gen_tcp:connect/3. Some options are forbidden, such as mode, packet, header, active. When tcp_opts is not provided, epgsql does some tuning (eg, sets TCP keepalive and auto-tunes buffer), but when tcp_opts is provided, no additional tweaks are added by epgsql itself, other than necessary ones (active, packet and mode).
async see Server notifications
codecs see Pluggable datatype codecs
nulls terms which will be used to represent SQL NULL. If any of those has been encountered in placeholder parameters ($1, $2 etc values), it will be interpreted as NULL. 1st element of the list will be used to represent NULLs received from the server. It's not recommended to use "string"s or lists. Try to keep this list short for performance! Default is [null, undefined], i.e. encode null or undefined in parameters as NULL and decode NULLs as atom null.
replication see Streaming replication protocol
application_name is an optional string parameter. It is usually set by an application upon connection to the server. The name will be displayed in the pg_stat_activity view and included in CSV log entries.
socket_active is an optional parameter, which can be true or an integer in the range -32768 to 32767 (inclusive, however only positive value make sense right now). This option is used to control the flow of incoming messages from the network socket to make sure huge query results won't result in epgsql process mailbox overflow. It affects the behaviour of some of the commands and interfaces (epgsqli and replication), so, use with caution! See Active socket for more details.

Options may be passed as proplist or as map with the same key names.

Asynchronous connect example (applies to epgsqli too):

  {ok, C} = epgsqla:start_link(),
  Ref = epgsqla:connect(C, "localhost", "username", "psss", #{database => "test_db"}),
  receive
    {C, Ref, connected} ->
        {ok, C};
    {C, Ref, Error = {error, _}} ->
        Error;
    {'EXIT', C, _Reason} ->
        {error, closed}
  end.

Simple Query

-include_lib("epgsql/include/epgsql.hrl").

-type query() :: string() | iodata().
-type squery_row() :: tuple() % tuple of binary().

-type ok_reply(RowType) ::
    {ok, ColumnsDescription :: [epgsql:column()], RowsValues :: [RowType]} |                            % select
    {ok, Count :: non_neg_integer()} |                                                            % update/insert/delete
    {ok, Count :: non_neg_integer(), ColumnsDescription :: [epgsql:column()], RowsValues :: [RowType]}. % update/insert/delete + returning
-type error_reply() :: {error, query_error()}.
-type reply(RowType) :: ok_reply() | error_reply().

-spec squery(connection(), query()) -> reply(squery_row()) | [reply(squery_row())].
%% @doc runs simple `SqlQuery' via given `Connection'
squery(Connection, SqlQuery) -> ...

examples:

epgsql:squery(C, "insert into account (name) values  ('alice'), ('bob')").
> {ok,2}

epgsql:squery(C, "select * from account").
> {ok,
    [#column{name = <<"id">>, type = int4, …},#column{name = <<"name">>, type = text, …}],
    [{<<"1">>,<<"alice">>},{<<"2">>,<<"bob">>}]
}

epgsql:squery(C,
    "insert into account(name)"
    "    values ('joe'), (null)"
    "    returning *").
> {ok,2,
    [#column{name = <<"id">>, type = int4, …}, #column{name = <<"name">>, type = text, …}],
    [{<<"3">>,<<"joe">>},{<<"4">>,null}]
}

epgsql:squery(C, "SELECT * FROM _nowhere_").
> {error,
   #error{severity = error,code = <<"42P01">>,
          codename = undefined_table,
          message = <<"relation \"_nowhere_\" does not exist">>,
          extra = [{file,<<"parse_relation.c">>},
                   {line,<<"1160">>},
                   {position,<<"15">>},
                   {routine,<<"parserOpenTable">>}]}}

The simple query protocol returns all columns as binary strings and does not support parameters binding.

Several queries separated by semicolon can be executed by squery.

[{ok, _, [{<<"1">>}]}, {ok, _, [{<<"2">>}]}] = epgsql:squery(C, "select 1; select 2").

epgsqla:squery/2 returns result as a single message:

  Ref = epgsqla:squery(C, Sql),
  receive
    {C, Ref, Result} -> Result
  end.

Result has the same format as return value of epgsql:squery/2.

epgsqli:squery/2 returns results incrementally for each query inside Sql and for each row:

Ref = epgsqli:squery(C, Sql),
receive
  {C, Ref, {columns, Columns}} ->
      %% columns description
      Columns;
  {C, Ref, {data, Row}} ->
      %% single data row
      Row;
  {C, Ref, {error, _E} = Error} ->
      Error;
  {C, Ref, {complete, {_Type, Count}}} ->
      %% execution of one insert/update/delete has finished
      {ok, Count}; % affected rows count
  {C, Ref, {complete, _Type}} ->
      %% execution of one select has finished
      ok;
  {C, Ref, done} ->
      %% execution of all queries from Sql has been finished
      done;
end.

Extended Query

{ok, Columns, Rows}        = epgsql:equery(C, "select ...", [Parameters]).
{ok, Count}                = epgsql:equery(C, "update ...", [Parameters]).
{ok, Count, Columns, Rows} = epgsql:equery(C, "insert ... returning ...", [Parameters]).
{error, Error}             = epgsql:equery(C, "invalid SQL", [Parameters]).

Parameters - optional list of values to be bound to $1, $2, $3, etc.

The extended query protocol combines parse, bind, and execute using the unnamed prepared statement and portal. A select statement returns {ok, Columns, Rows}, insert/update/delete returns {ok, Count} or {ok, Count, Columns, Rows} when a returning clause is present. When an error occurs, all statements result in {error, #error{}}.

epgsql:equery(C, "select id from account where name = $1", ["alice"]),
> {ok,
    [#column{name = <<"id">>, type = int4, …}],
    [{1}]
}

PostgreSQL's binary format is used to return integers as Erlang integers, floats as floats, bytes/text/varchar columns as binaries, bools as true/false, etc. For details see pgsql_binary.erl and the Data Representation section below.

Asynchronous API epgsqla:equery/3 requires you to parse statement beforehand

#statement{types = Types} = Statement,
TypedParameters = lists:zip(Types, Parameters),
Ref = epgsqla:equery(C, Statement, [TypedParameters]),
receive
  {C, Ref, Res} -> Res
end.

Statement - parsed statement (see parse below)
Res has same format as return value of epgsql:equery/3.

epgsqli:equery(C, Statement, [TypedParameters]) sends same set of messages as squery including final {C, Ref, done}.

Prepared Query

{ok, Columns, Rows}        = epgsql:prepared_query(C, Statement :: #statement{} | string(), [Parameters]).
{ok, Count}                = epgsql:prepared_query(C, Statement, [Parameters]).
{ok, Count, Columns, Rows} = epgsql:prepared_query(C, Statement, [Parameters]).
{error, Error}             = epgsql:prepared_query(C, "non_existent_query", [Parameters]).

Parameters - optional list of values to be bound to $1, $2, $3, etc.
Statement - name of query given with erlang epgsql:parse(C, StatementName, "select ...", []). (can be empty string) or #statement{} record returned by epgsql:parse.

With prepared query one can parse a query giving it a name with epgsql:parse on start and reuse the name for all further queries with different parameters.

{ok, Stmt} = epgsql:parse(C, "inc", "select $1+1", []).
epgsql:prepared_query(C, Stmt, [4]).
epgsql:prepared_query(C, Stmt, [1]).

Asynchronous API epgsqla:prepared_query/3 requires you to always parse statement beforehand

#statement{types = Types} = Statement,
TypedParameters = lists:zip(Types, Parameters),
Ref = epgsqla:prepared_query(C, Statement, [TypedParameters]),
receive
  {C, Ref, Res} -> Res
end.

Statement - parsed statement (see parse below)
Res has same format as return value of epgsql:prepared_query/3.

epgsqli:prepared_query(C, Statement, [TypedParameters]) sends same set of messages as squery including final {C, Ref, done}.

Parse/Bind/Execute

{ok, Statement} = epgsql:parse(C, [StatementName], Sql, [ParameterTypes]).

StatementName - optional, reusable, name for the prepared statement.
ParameterTypes - optional list of PostgreSQL types for each parameter.

For valid type names see pgsql_types.erl.

epgsqla:parse/2 sends {C, Ref, {ok, Statement} | {error, Reason}}.

epgsqli:parse/2 sends:

{C, Ref, {types, Types}}
{C, Ref, {columns, Columns}}
{C, Ref, no_data} if statement will not return rows
{C, Ref, {error, Reason}}

ok = epgsql:bind(C, Statement, [PortalName], ParameterValues).

PortalName - optional name for the result portal.

both epgsqla:bind/3 and epgsqli:bind/3 send {C, Ref, ok | {error, Reason}}

{ok | partial, Rows} = epgsql:execute(C, Statement, [PortalName], [MaxRows]).
{ok, Count}          = epgsql:execute(C, Statement, [PortalName]).
{ok, Count, Rows}    = epgsql:execute(C, Statement, [PortalName]).

PortalName - optional portal name used in epgsql:bind/4.
MaxRows - maximum number of rows to return (0 for all rows).

epgsql:execute/3 returns {partial, Rows} when more rows are available.

epgsqla:execute/3 sends {C, Ref, Result} where Result has same format as return value of epgsql:execute/3.

epgsqli:execute/3 sends

{C, Ref, {data, Row}}
{C, Ref, {error, Reason}}
{C, Ref, suspended} partial result was sent, more rows are available
{C, Ref, {complete, {_Type, Count}}}
{C, Ref, {complete, _Type}}

ok = epgsql:close(C, Statement).
ok = epgsql:close(C, statement | portal, Name).
ok = epgsql:sync(C).

All epgsql functions return {error, Error} when an error occurs.

epgsqla/epgsqli modules' close and sync functions send {C, Ref, ok}.

Batch execution

Batch execution is bind + execute for several prepared statements. It uses unnamed portals and MaxRows = 0.

Results = epgsql:execute_batch(C, BatchStmt :: [{statement(), [bind_param()]}]).
{Columns, Results} = epgsql:execute_batch(C, statement() | sql_query(), Batch :: [ [bind_param()] ]).

BatchStmt - list of {Statement, ParameterValues}, each item has it's own #statement{}
Batch - list of ParameterValues, each item executes the same common #statement{} or SQL query
Columns - list of #column{} descriptions of Results columns
Results - list of {ok, Count} or {ok, Count, Rows}

There are 2 versions:

execute_batch/2 - each item in a batch has it's own named statement (but it's allowed to have duplicates)

example:

{ok, S1} = epgsql:parse(C, "one", "select $1::integer", []),
{ok, S2} = epgsql:parse(C, "two", "select $1::integer + $2::integer", []),
[{ok, [{1}]}, {ok, [{3}]}] = epgsql:execute_batch(C, [{S1, [1]}, {S2, [1, 2]}]).
ok = epgsql:close(C, "one").
ok = epgsql:close(C, "two").

execute_batch/3 - each item in a batch executed with the same common SQL query or #statement{}. It's allowed to use unnamed statement.

example (the most efficient way to make batch inserts with epgsql):

{ok, Stmt} = epgsql:parse(C, "my_insert", "INSERT INTO account (name, age) VALUES ($1, $2) RETURNING id", []).
{[#column{name = <<"id">>}], [{ok, [{1}]}, {ok, [{2}]}, {ok, [{3}]}]} =
    epgsql:execute_batch(C, Stmt, [ ["Joe", 35], ["Paul", 26], ["Mary", 24] ]).
ok = epgsql:close(C, "my_insert").

equivalent:

epgsql:execute_batch(C, "INSERT INTO account (name, age) VALUES ($1, $2) RETURNING id",
                     [ ["Joe", 35], ["Paul", 26], ["Mary", 24] ]).

In case one of the batch items causes an error, all the remaining queries of that batch will be ignored. So, last element of the result list will be {error, #error{}} and the length of the result list might be shorter that the length of the batch. For a better illustration of such scenario please refer to epgsql_SUITE:batch_error/1

epgsqla:execute_batch/{2,3} sends {C, Ref, Results}

epgsqli:execute_batch/{2,3} sends

{C, Ref, {data, Row}}
{C, Ref, {error, Reason}}
{C, Ref, {complete, {_Type, Count}}}
{C, Ref, {complete, _Type}}
{C, Ref, done} - execution of all queries from Batch has finished

Query cancellation

epgsql:cancel(connection()) -> ok.

PostgreSQL protocol supports cancellation of currently executing command. cancel/1 sends a cancellation request via the new temporary TCP/TLS_over_TCP connection asynchronously, it doesn't await for the command to be cancelled. Instead, client should expect to get {error, #error{code = <<"57014">>, codename = query_canceled}} back from the command that was cancelled. However, normal response can still be received as well. While it's not so straightforward to use with synchronous epgsql API, it plays quite nicely with asynchronous epgsqla API. For example, that's how a query with soft timeout can be implemented:

squery(C, SQL, Timeout) ->
    Ref = epgsqla:squery(C, SQL),
    receive
       {C, Ref, Result} -> Result
    after Timeout ->
        ok = epgsql:cancel(C),
        % We can still receive {ok, …} as well as
        % {error, #error{codename = query_canceled}} or any other error
        receive
            {C, Ref, Result} -> Result
        end
    end.

This API should be used with extreme care when pipelining is in use: it only cancels currently executing command, all the subsequent pipelined commands will continue their normal execution. And it's not always easy to see which command exactly is executing when we are issuing the cancellation request.

Data Representation

Data representation may be configured using pluggable datatype codecs, so following is just default mapping:

PG type	Representation
null	`null`
bool	`true`
char	`$A`
intX	`1`
floatX	`1.0`
date	`{Year, Month, Day}`
time	`{Hour, Minute, Second.Microsecond}`
timetz	`{time, Timezone}`
timestamp	`{date, time}`
timestamptz	`{date, time}`
interval	`{time, Days, Months}`
text	`<<"a">>`
varchar	`<<"a">>`
bytea	`<<1, 2>>`
array	`[1, 2, 3]`
record	`{int2, time, text, ...}` (decode only)
point	`{10.2, 100.12}`
int4range	`[1,5)`
hstore	`{[ {binary(), binary() \| null} ]}` (configurable)
json/jsonb	`<<"{ \"key\": [ 1, 1.0, true, \"string\" ] }">>` (configurable)
uuid	`<<"123e4567-e89b-12d3-a456-426655440000">>`
inet	`inet:ip_address()`
cidr	`{ip_address(), Mask :: 0..32}`
macaddr(8)	tuple of 6 or 8 `byte()`
geometry	`ewkb:geometry()`
tsrange	`{{Hour, Minute, Second.Microsecond}, {Hour, Minute, Second.Microsecond}}`
tstzrange	`{{Hour, Minute, Second.Microsecond}, {Hour, Minute, Second.Microsecond}}`
daterange	`{{Year, Month, Day}, {Year, Month, Day}}`

null can be configured. See nulls connect/1 option.

timestamp and timestamptz parameters can take erlang:now() format: {MegaSeconds, Seconds, MicroSeconds}

int4range is a range type for ints that obeys inclusive/exclusive semantics, bracket and parentheses respectively. Additionally, infinities are represented by the atoms minus_infinity and plus_infinity

tsrange, tstzrange, daterange are range types for timestamp, timestamptz and date respectively. They can return empty atom as the result from a database if bounds are equal

hstore type can take map or jiffy-style objects as input. Output can be tuned by providing return :: map | jiffy | proplist option to choose the format to which hstore should be decoded. nulls :: [atom(), ...] option can be used to select the terms which should be interpreted as SQL NULL - semantics is the same as for connect/1 nulls option.

json and jsonb types can optionally use a custom JSON encoding/decoding module to accept and return erlang-formatted JSON. The module must implement the callbacks in epgsql_codec_json, which most popular open source JSON parsers will already, and you can specify it in the codec configuration like this:

{epgsql_codec_json, JsonMod}

% With options
{epgsql_codec_json, JsonMod, EncodeOpts, DecodeOpts}

% Real world example using jiffy to return a map on decode
{epgsql_codec_json, {jiffy, [], [return_maps]}}

Note that the decoded terms will be message-passed to the receiving process (i.e. copied), which may exhibit a performance hit if decoding large terms very frequently.

Errors

Errors originating from the PostgreSQL backend are returned as {error, #error{}}, see epgsql.hrl for the record definition. epgsql functions may also return {error, What} where What is one of the following:

{unsupported_auth_method, Method} - required auth method is unsupported
timeout - request timed out
closed - connection was closed
sync_required - error occurred and epgsql:sync must be called

Server Notifications

PostgreSQL may deliver two types of asynchronous message: "notices" in response to notice and warning messages generated by the server, and notifications which are generated by the LISTEN/NOTIFY mechanism.

Passing the {async, PidOrName} option to epgsql:connect/3 will result in these async messages being sent to the specified pid or registered process, otherwise they will be dropped.

Another way to set notification receiver is to use set_notice_receiver/2 function. It returns previous async value. Use undefined to disable notifications.

% receiver is pid()
{ok, Previous} = epgsql:set_notice_receiver(C, self()).

% receiver is registered process
register(notify_receiver, self()).
{ok, Previous1} = epgsqla:set_notice_receiver(C, notify_receiver).

% disable notifications
{ok, Previous2} = epgsqli:set_notice_receiver(C, undefined).

Message formats:

{epgsql, Connection, {notification, Channel, Pid, Payload}}

Connection - connection the notification occurred on
Channel - channel the notification occurred on
Pid - database session pid that sent notification
Payload - optional payload, only available from PostgreSQL >= 9.0

{epgsql, Connection, {notice, Error}}

Connection - connection the notice occurred on
Error - an #error{} record, see epgsql.hrl

Utility functions

Transaction helpers

with_transaction(connection(), fun((connection()) -> Result :: any()), Opts) ->
    Result | {rollback, Reason :: any()} when
Opts :: [{reraise, boolean()},
         {ensure_committed, boolean()},
         {begin_opts, iodata()}] | map().

Executes a function in a PostgreSQL transaction. It executes BEGIN prior to executing the function, ROLLBACK if the function raises an exception and COMMIT if the function returns without an error. If it is successful, it returns the result of the function. The failure case may differ, depending on the options passed.

Options (proplist or map):

reraise (default true): when set to true, the original exception will be re-thrown after rollback, otherwise {rollback, ErrorReason} will be returned
ensure_committed (default false): even when the callback returns without exception, check that transaction was committed by checking the CommandComplete status of the COMMIT command. If the transaction was rolled back, the status will be rollback instead of commit and an ensure_committed_failed error will be generated.
begin_opts (default ""): append extra options to BEGIN command (see https://www.postgresql.org/docs/current/static/sql-begin.html) as a string by just appending them to "BEGIN " string. Eg {begin_opts, "ISOLATION LEVEL SERIALIZABLE"}. Beware of SQL injection! The value of begin_opts is not escaped!

Command status

epgsql{a,i}:get_cmd_status(C) -> undefined | atom() | {atom(), integer()}

This function returns the last executed command's status information. It's usually the name of SQL command and, for some of them (like UPDATE or INSERT) the number of affected rows. See libpq PQcmdStatus. But there is one interesting case: if you execute COMMIT on a failed transaction, status will be rollback, not commit. This is how you can detect failed transactions:

{ok, _, _} = epgsql:squery(C, "BEGIN").
{error, _} = epgsql:equery(C, "SELECT 1 / $1::integer", [0]).
{ok, _, _} = epgsql:squery(C, "COMMIT").
{ok, rollback} = epgsql:get_cmd_status(C).

Server parameters

epgsql{a,i}:get_parameter(C, Name) -> binary() | undefined

Retrieve actual value of server-side parameters, such as character endoding, date/time format and timezone, server version and so on. See libpq PQparameterStatus. Parameter's value may change during connection's lifetime.

Active socket

By default epgsql sets its underlying gen_tcp or ssl socket into {active, true} mode (make sure you understand the OTP inet:setopts/2 documentation about active option). That means if PostgreSQL decides to quickly send a huge amount of data to the client (for example, client made a SELECT that returns large amount of results or when we are connected in streaming replication mode and receiving a lot of updates), underlying network socket might quickly send large number of messages to the epgsql connection process leading to the growing mailbox and high RAM consumption (or even OOM situation in case of really large query result or massive replication update).

To avoid such scenarios, epgsql can may rely on "TCP backpressure" to prevent socket from sending unlimited number of messages - implement a "flow control". To do so, socket_active => 1..32767 could be added at connection time. This option would set {active, N} option on the underlying socket and would tell the network to send no more than N messages to epgsql connection and then pause to let epgsql and the client process the already received network data and then decide how to proceed.

The way this pause is signalled to the client and how the socket can be activated again depends on the interface client is using:

when epgsqli interface is used, epgsql would send all the normal low level messages and then at any point it may send {epgsql, C, socket_passive} message to signal that socket have been paused. epgsql:activate(C) must be called to re-activate the socket.
when epgsql is connected in Streaming replication mode and pid() is used as the receiver of the X-Log Data messages, it would behave in the same way: {epgsql, C, socket_passive} might be sent along with {epgsql, self(), {x_log_data, _, _, _}} messages and epgsql:activate/1 can be used to re-activate.
in all the other cases (epgsql / epgsqla command, while COPY FROM STDIN mode is active, when Streaming replication with Erlang module callback as receiver of X-Log Data or while connection is idle) epgsql would transparently re-activate the socket automatically: it won't prevent high RAM usage from large SELECT result, but it would make sure epgsql process has no more than N messages from the network in its mailbox.

It is a good idea to combine socket_active => N with some specific value of tcp_opts => [{buffer, X}] since each of the N messages sent from the network to epgsql process would contain no more than X bytes. So the MAXIMUM amount of data seating at the epgsql mailbox could be roughly estimated as N * X. So if N = 256 and X = 512*1024 (512kb) then there will be no more than N * X = 256 * 524288 = 134_217_728 or 128MB of data in the mailbox at the same time.