Skip to content

conformance: the hangs this page calls the largest debt are 75 -> 8, and the real largest debt read as zero - #51

Merged
tannevaled merged 1 commit into
mainfrom
conformance-largest-debt
Nov 15, 2026
Merged

tannevaled merged 1 commit into
mainfrom
conformance-largest-debt

Conversation

@tannevaled

Copy link
Copy Markdown
Contributor

conformance: the hangs this page calls the largest debt are 75 → 8, and the real largest debt read as zero

The page's second "reading that matters" was "The largest single debt is a
hang, not a missing feature"
— Net::FTP 442 examples behind, 39 of its 56
spec files killed by the clock. Both halves of that sentence are now false,
and it is the one a reader carries away.

The hangs: 75 → 8

Re-measured 2026-11-13 against MRI 4.0.6, same corpus, 30 s per file. The page
now carries this as a then/now table with the cause in each row, because the
causes are the part that transfers:

then now the cause
library/net-ftp 39 killed, −442 0 killed, −128 a blocking socket call held the emulated GVL — on the data connection
library/net-http 27 killed 0 TCPSocket.open { … } never closed the socket
library/socket 8 killed 7 Thread#kill could not interrupt a blocking native call
library/timeout 1 killed 1 the same, twelve sites left

One symptom, three unrelated causes — which is the reason the first fix
looked like it had done nothing: it was measured against the specs belonging to
the other half of the same mechanism. The net-ftp row is the sharp case. The
control connection was bounded and released the GVL; dialData, acceptData,
sendData, retrlines and retrbinary still held it, and in active mode —
which ruby/spec's fixture selects — holding the GVL inside that Accept parks the
very Ruby thread that would connect.

The real largest debt, which read as zero

library/socket's own spec_helper.rb calls MSpec.enable_feature, which the
conformance shim did not define, so all 190 of those files raised NameError at
load and scored 0 — including under MRI.

A control of zero is indistinguishable from a module with no gap. The same sweep
that now reads −1566 read "+217 in rbgo's favour" twenty minutes earlier.

Repaired and measured 2026-11-14 against MRI 4.0.6: MRI 1790 / rbgo 224, so
library/socket alone is about −1560 — larger than every module in "the rest
is concentrated"
put together. It is unimplemented surface, not hangs:
Addrinfo#initialize −151, BasicSocket#recvmsg + #recvmsg_nonblock −100, the
address and option accessors −188 together. That list is now marked as having
been compiled while socket contributed a flat zero on both sides.

⛔ The table's gap is relabelled a floor, not a total, and the two figures are
deliberately not added together: different runs, different MRI patch releases
(4.0.7 and 4.0.6), and a sum across two measurements is how two errors cancel.

Checks

  • go run ./tools/conformancefigure --root . passes — 5 total claims, 2 version
    claims and 6 stubbed-helper mentions checked against the gate's 23 672 at Ruby
    4.0.9.
  • go test ./... green across all six packages.
  • No Go files touched; the change is one content page.

The gates had nothing to say about any figure corrected here, and that is their
honest limit rather than a fault: the dated run that produced these numbers is
gone, so there is nothing to re-derive them from. What catches this is someone
asking which published sentences the merges made false.

Refs go-embedded-ruby/ruby#916, go-embedded-ruby/ruby#933, go-embedded-ruby/ruby#936, go-embedded-ruby/ruby#937, go-embedded-ruby/ruby#938

(All stay open or are handled in their own repository. The matching corrections
to the rbgo README and to the org landing page are separate pull requests — one
surface each.)

🤖 Generated with Claude Code

…and the real largest debt read as zero

The page's second "reading that matters" was **"The largest single debt is a
hang, not a missing feature"** -- Net::FTP 442 examples behind, 39 of its 56
files killed by the clock. Both halves of that sentence are now false, and it is
the one a reader would carry away.

**The hangs.** Re-measured 2026-11-13 against MRI 4.0.6, same corpus, 30 s per
file: net-ftp 0 killed and -128, net-http 0 from 27, socket 7 from 8, timeout 1
-- 75 across the corpus down to 8. One symptom, three unrelated causes: a
blocking socket call holding rbgo's emulated GVL on the *data* connection (which
is why fixing the control connection first moved nothing at all), a
`TCPSocket.open { }` that never closed the socket, and a `Thread#kill` unable to
interrupt a blocking native call. The page now carries that as a then/now table
with the cause in each row, because the causes are the part that transfers.

**The real largest debt.** `library/socket`'s own `spec_helper.rb` calls
`MSpec.enable_feature`, which the shim did not define, so all 190 of those files
raised `NameError` at load and scored 0 -- including under MRI. A control of zero
is indistinguishable from a module with no gap. Repaired, the module is MRI 1790
/ rbgo 224: about **-1560**, larger than every module in "the rest is
concentrated" put together, and it is unimplemented surface rather than hangs.
That list is now marked as having been compiled while socket contributed zero.

The table's gap is relabelled a **floor**. The two totals are not added up here:
different runs, different MRI patch releases (4.0.7 and 4.0.6), and a sum across
two measurements is how two errors cancel.

Checks: `go run ./tools/conformancefigure --root .` passes (5 total claims, 2
version claims, 6 stubbed-helper mentions against 23672 at Ruby 4.0.9), and
`go test ./...` is green across all six packages. No Go files are touched.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@tannevaled
tannevaled merged commit c4e0ce0 into main Nov 15, 2026
2 checks passed
@tannevaled
tannevaled deleted the conformance-largest-debt branch November 15, 2026 09:39
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant