Net::FTP has no #initialize: construction is wired to .new and to nothing else (-38 examples, the largest item left in net-ftp)

Net::FTP has no `#initialize`: construction is wired to `.new` and to nothing else

Measured against MRI 4.0.6 (net-ftp 0.3.9) over all 56 library/net-ftp files,
same corpus, 30 s per file:

engine	passing
MRI 4.0.6	459
rbgo	331
	−128

initialize_spec.rb alone is −38 of that, and it is the largest single item
left in the module: MRI passes 41, rbgo passes 3.

The cause

Net::FTP#initialize is not defined. Construction lives entirely in the
class method Net::FTP.new, so the class inherits Object#initialize, which
takes no arguments:

RBGO_RESULT pass=3 fail=2 error=36 skip=0
RBGO_DETAIL error ArgumentError  Net::FTP#initialize when passed host | tries to connect to the passed host
                                 wrong number of arguments (given 1, expected 0)

ruby/spec drives initialize directly, which is normal for this module —
it is how the suite tests construction without connecting:

@ftp.send(:initialize, nil, options)
@ftp.passive.should == false

and it asserts the method is there and private:

it "is private" do
  Net::FTP.private_instance_methods(false).should.include?(:initialize)
end

The shape

This is the defect this repository keeps producing: a mechanism wired at only
some of its doors. Construction works perfectly through .new and does not
exist through #initialize, which is the door MRI documents as a private
instance method and which ruby/spec uses throughout.

It is the tenth instance recorded this week, after ioBlock in Net::HTTP but
not Net::FTP, read_timeout read from the options Hash but not
open_timeout, and ftpInit returning before the options when the host was
nil.

Shape of the fix

Define Net::FTP#initialize(host = nil, options = {}) as a private instance
method carrying what ftpInit does today, and make Net::FTP.new allocate
and call it. Then .new, .open and send(:initialize, …) all go through one
implementation instead of one door being a special case.

Worth checking while there: MRI calls options.to_hash on the argument, so an
object that converts counts as an options Hash; rbgo currently requires a
literal Hash.

Not a hang

For the record, since this module is cited in #916: with a 30 s per-file limit,
no file is killed by the clock under either engine — specdelta prints no
KILLED on timeout row in this sweep. The remaining −128 is ordinary
unimplemented behaviour, not the hang symptom.

Prior art: priorart "Net::FTP#initialize private instance method send" at 2026-11-13 14:40 CET reported prior art; checked: priorart --repo go-embedded-ruby/ruby 'Net::FTP#initialize private instance method send' returned 508 broad keyword matches; the nearest are #916 (the net-ftp hang issue, which this explicitly is NOT), #928 (my own open PR, which adds default_passive and fixes ftpInit dropping options on a nil host but does not define #initialize), and #621 (an mspec shim gap). None covers Net::FTP#initialize being undefined.

Before implementing this: `Net::FTP.allocate` does not produce a `Net::FTP`

Defining #initialize is necessary and not sufficient, because of how the
spec reaches it:

before :each do
  @ftp = Net::FTP.allocate
  @ftp.stub!(:connect)
end

Measured on main:

$ rbgo -e 'require "net/ftp"; a = Net::FTP.allocate; p a.class; p a.passive'
Net::FTP
ArgumentError: passive: interface conversion: object.Value is *vm.RObject, not *vm.ftpObj

allocate hands back a generic *RObject that merely reports the class
Net::FTP. Every accessor then fails its type assertion, so an #initialize
defined on top of it would be handed an object it cannot populate. The 38
examples in initialize_spec.rb all start from that allocate.

So the work is larger than "define a private #initialize":

a class whose instances are backed by a Go struct needs an allocator the
class itself provides, and Class#allocate has to consult it instead of
building a plain RObject;
then #initialize can be a private instance method carrying what ftpInit
does today, with .new becoming allocate-then-initialize;
and only then does send(:initialize, nil, options) reach a real ftpObj.

Step 1 is not specific to Net::FTP — every native-backed class has the same
hole, so it wants doing once in the object model rather than per binding. That
is a different and wider change than this issue was filed for, and worth saying
before someone starts at step 2.

⛔ A separate, smaller defect found on the way

The failure above is a Go interface-conversion message reaching Ruby:

ArgumentError: passive: interface conversion: object.Value is *vm.RObject, not *vm.ftpObj

Internal type names have no business in a user-visible exception, and the class
is wrong too — MRI would raise NoMethodError or TypeError here, not
ArgumentError. Whatever happens to the rest of this issue, that message is
worth not shipping: it tells an embedder about *vm.ftpObj, and it will be
produced by any native-backed class reached through allocate.

Navigation Menu

Repository navigation

Net::FTP has no #initialize: construction is wired to .new and to nothing else (-38 examples, the largest item left in net-ftp) #929