Jens Geyer created THRIFT-6279:
----------------------------------
Summary: Erlang: the JSON protocol cannot read a message and
writes binary strings unquoted
Key: THRIFT-6279
URL: https://issues.apache.org/jira/browse/THRIFT-6279
Project: Thrift
Issue Type: Bug
Components: Erlang - Library
Reporter: Jens Geyer
Found while working on THRIFT-6269. {{thrift_json_protocol}} does not
round-trip a message. The Erlang cross test covers only binary and compact, and
{{lib/erl/test}} has no JSON test, so nothing notices. Measured on master with
OTP 25 and 28.
h2. 1. Reading any message fails
{{convert_data(integer, I)}} calls {{list_to_integer(I)}}. That dates from
THRIFT-1227, when the protocol used jsx. THRIFT-1357 (2014) replaced jsx with
{{thrift_json_parser}}, whose {{format_number/1}} already returns integers.
Since then {{list_to_integer/1}} is handed an integer.
* A message the protocol wrote itself fails to read at its version number with
{{badarg}}.
* The integer clause of {{read(_, byte)}} has the same call.
h2. 2. A string given as a binary is written without quotes
The generator's default {{string=both}} accepts a string field as a binary, but
{{write(_, \{string, Str\})}} writes a binary unchanged.
* Measured: the field comes out as {{"str":abc}}, which is not JSON.
* A string given as a list is quoted, but nothing in it is escaped.
h2. 3. Over a socket the read never ends
{{read_all/1}} reads until the transport returns nothing more, which a socket
or framed transport never does: it waits for more data.
Measured with {{thrift_socket_server}} and {{thrift_client_util}} using the
JSON protocol, both buffered and framed:
* The server process dies at its 500 ms receive timeout with a {{case_clause}}
on {{\{error,timeout\}}}.
* The client call returns an error with {{case_clause}} on {{\{error,closed\}}}.
* The binary protocol works in the same setup.
h2. 4. One transport call per byte
{{read_all_1/2}} asks the transport for one byte at a time. This is the second
point of THRIFT-6269, left out of PR #3867 because of 1 to 3.
h2. Suggested
* Fix 1 and 2, with round-trip tests over a memory buffer. For the list case,
the test should include characters that need escaping.
* Make the read end with the message, not with the stream: parse incrementally,
or read what the parser still needs.
* Then read in larger pieces, and resolve the {{\{error, 'EOF'\}}} clause in
{{read_all_1/2}}. Its only producer, {{thrift_http_transport:read/2}}, cannot
reach it.
* Add JSON to the Erlang entry of {{test/tests.json}} once it works.
_Drafted with AI assistance (Claude Opus 5); reviewed and filed by Jens Geyer._
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
