test_live_full_cycle fails in CI (400 collision or job never finishes) — blocks all PRs

[VasilevNStas]

Problem

test_live_full_cycle in test/test_baza_live.rb:36-62 periodically fails with:

BazaRb::ServerFailure: Invalid response code #400 at PUT https://api.zerocracy.com/push/fakeea9e3b3fea801116
(Flash: An existing job named "fakeea9e3b3fea801116" is running now)

This happens in CI of every open PR: #350, #351, #352, #310, #317, #314, #308, and others.

Root cause

The job name is generated at line 133-134:

def fake_name
  "fake#{SecureRandom.hex(8)}"
end

SecureRandom.hex(8) produces 16 hex chars — 2^64 possible values. When CI runs in parallel (macOS + ubuntu create jobs simultaneously) or when a previous run left a job with the same name still running, the zerocracy server returns 400.

The test fails not because the code is broken, but because randomness does not guarantee uniqueness. "Re-run CI and hope it passes" is the only workaround — a luck-based technique, not determinism.

Proposed fixes

Option A (recommended): add timestamp

def fake_name
  "fake-#{Time.now.to_i}-#{SecureRandom.hex(4)}"
end

The name now contains the creation timestamp + 8 hex chars. Even if two CI jobs start at the same second, hex(4) still gives 4 billion variants. Collision becomes practically impossible.

Additionally: extract the push call from test_live_full_cycle into a separate setup method that runs once before all live tests, to reduce server-side garbage.

Option B: UUID

def fake_name
  "fake-#{SecureRandom.uuid}"
end

UUID provides 122 bits of entropy. Collision within a single CI run is impossible.

Option C: retry on 400

def fake_name
  loop do
    n = "fake#{SecureRandom.hex(8)}"
    return n unless LIVE.name_exists?(n)
  end
end

But this adds an extra HTTP request and does not fix the case when the job exists but has not been created yet.

Expected outcome

After the fix, test_live_full_cycle should pass 100% of runs without requiring re-runs.

[VasilevNStas]

@yegor256 please add the "help wanted" label to this issue

[VasilevNStas]

Update: additional failure mode

Since the issue was filed, the CI failure signature has changed. The test no longer fails with a 400 collision — it now fails with:

Job #75903 ("fakeaee4ce5329be0126") did not finish in 0s at api.zerocracy.com,
which most probably means it got stuck on the Zerocracy platform and never completed

This indicates the job IS created successfully (no 400), but the platform never completes execution within the 2-minute wait_for timeout.

Root cause (current)

The Zerocracy platform's job runner appears to be stuck or overloaded — jobs are accepted but never transition to finished state. This is a server-side infrastructure issue.

Impact

Every open PR on the repository fails CI because test_live_full_cycle is part of the default rake test suite. As of 2026-06-05, the following PRs are affected (all created by @VasilevNStas):

• #360: extract Compress module (unzip/zipped) from god-class #392 (god-class compress extract)
• #361: cache CSRF token with 10 min TTL #391 (CSRF caching)
• #368/#369: shared Validation module + contract test #390 (Validation module)
• #362: add BazaRb::Error hierarchy for custom errors #389 (Error hierarchy)
• #303: Fix enter re-yielding user block on transient POST failure #388 (enter re-yield fix)
• #341: Fix retry count inconsistency - attempt was one short of await/recover #387 (retry count fix)
• #349: Fix IO leak and retry with empty body in durable_place #386 (IO leak fix)
• #355: Make rehost opt-in to prevent SSRF via X-Zerocracy-Host #384 (SSRF fix)
• #356: Protect unzip against compression bombs #382 (gzip bomb fix)
• #358: Allow backoff to be configured via BAZA_BACKOFF env var #381 (backoff env var)
• #372: Add exponential backoff for blanks sleep in download #380 (blanks backoff fix)
• #371: Extract backoff calculation into private helper #379 (backoff helper)
• #370: Merge await and recover into single retry_on method #378 (retry_on merge)
• #367: Add pname format and length validation to durable_place/durable_find #377 (pname validation)
• #366: Add pname format and length validation to lock/unlock/name_exists?/recent #376 (enter guards)
• #365: Add pname format and length validation to BazaRb#push #375 (validate module)
• #363: BigDecimal precision for amount formatting in transfer/fee #374 (BigDecimal)
• #364: use Base64.strict_encode64 instead of encode64 + delete newline #373 (strict_encode64)
• #345: Fix Content-Type application/zip on gzip-encoded upload bodies #350 (Content-Type fix)

None of these PRs modify test/test_baza_live.rb or any code that could affect the live test.

Additional bug in the test

Line 52 of test/test_baza_live.rb has a bug in the error message:

"Job ##{id} (#{n.inspect}) did not finish in #{(Time.now - Time.now).round}s at #{HOST}"

Time.now - Time.now always evaluates to 0, so the message always shows "0s" instead of the actual elapsed time. It should capture start = Time.now before wait_for and use (Time.now - start).round in the message.

Proposed solutions

1. Server-side: Zerocracy maintainers need to investigate why jobs are not completing on the platform
2. Test-side: Increase the wait_for timeout from 2 to 5 minutes, or skip test_live_full_cycle in CI runs (only run manually)
3. Fix the bug: Correct the Time.now - Time.now error message to show actual elapsed time

Workaround

Until this is resolved, all PRs show a false-positive CI failure. Reviewers must check that only test_live_full_cycle fails and all other 118 tests pass.