odba 1.2.2 → 1.2.3

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: e275634934d76a06600ef71830600a6a6cd1e03c3f59069aa94f998f25d9a973
-  data.tar.gz: cf0ae8760286cafa23e4d1c1c37e304aa8a548daa8438c00e58c1f2492691743
+  metadata.gz: a5870eb15332e2b3e862dd27adf014e175e46588d673527794988a3a876f61b7
+  data.tar.gz: 8209e0f9122128f21e004f9db3a2fc7b58686f1c2b6f802ce3be628d095cd343
 SHA512:
-  metadata.gz: 9d7005efee025dddcf75ae2e80b50266482a322916b59a7bd25fa339149a5df762e87248d4351e80acf819550ab2738d2f3cb3d04046c759bddfb48418221b69
-  data.tar.gz: cd31af260988120f4ab9c4b73126896f8c0ce43d68b9b4da24ffd1e968fc8fd0c3564fec466ba5f1d2ecd36c5605afe2b3b124bcd5ef2eb990b7535bc29e3751
+  metadata.gz: e7a633a8aeecdd7d243bce697800a3163b9090cae5352609aac8139704594bb02e71521f443829efeb618e1882c7f94b19b014c03b4e606c62ba66832566c173
+  data.tar.gz: abb1a74dff537de863797b397185e05ef1c6301da2d0e51ffcd48b3621b6a89f5933534b4fca02067d305abcb5645988b8eb805a07432d4e9e358fbf9d2cf7ea

data/History.md

@@ -1,3 +1,25 @@
+## 1.2.3 / 01.09.2026
+
+1.2.2 narrowed the rescue around the peer notification in `Cache#next_id` too
+far. "A peer we cannot reach" is not only `DRb::DRbError`: a reference into a
+peer that has restarted, or whose entry has expired from the DRb object space,
+raises `RangeError("invalid reference")` from `DRbObjectSpace#to_obj` - a
+StandardError, and not a DRbError. 1.1.9's classless rescue absorbed it; 1.2.2
+let it out and it reached the caller.
+
+In oddb.org that showed up on the first night under 1.2.2: three index
+rebuilds died with `invalid reference (druby://127.0.0.1:10000)`. An index
+that is not built is a deferred index, and ODBA fills a deferred index in
+`Cache#setup` - so the *next* process to start died on it too, before doing
+any work of its own.
+
+* `next_id` re-raises `OdbaDuplicateIdError` (the one exception that must
+  reach the retry, which is what 1.2.2 was for) and swallows every other
+  StandardError from a peer, as 1.1.9 did.
+* Two regression tests, one per direction: a stale reference must not stop
+  the allocation, and the catch-all must not swallow the conflict again.
+  The first fails against 1.2.2 with the production error.
+
 ## 1.2.2 / 31.08.2026
 
 Two processes on one database handed out the same odba_id. Whichever wrote

data/lib/odba/cache.rb

@@ -441,14 +441,20 @@ module ODBA
       end
       @peers.each do |peer|
         peer.reserve_next_id id
-      rescue DRb::DRbError
-        # A peer we cannot reach must not stop the allocation. Note the
-        # explicit class: a bare rescue here also swallowed the
-        # OdbaDuplicateIdError a peer raises when the id is already taken,
-        # so the retry below could never run and both processes kept the
-        # same id. Whichever wrote last overwrote the other's row in
-        # `object`, and every reference to it then resolved to a foreign
-        # object.
+      rescue OdbaDuplicateIdError
+        # The one exception that must get through, to the retry below. Until
+        # 1.2.2 a bare rescue swallowed it, so the retry could never run and
+        # both processes kept the same id; whichever wrote last overwrote the
+        # other's row in `object`, and every reference to it then resolved to
+        # a foreign object.
+        raise
+      rescue StandardError
+        # Anything else means we could not reach this peer, and that must not
+        # stop the allocation. Naming DRb::DRbError alone is not enough, which
+        # 1.2.2 got wrong: a stale reference into a peer raises RangeError
+        # ("invalid reference", drb.rb DRbObjectSpace#to_obj), which is not a
+        # DRbError. On 01.09.2026 that took down three index rebuilds in
+        # oddb.org - the peer had simply gone away.
       end
       id
     rescue OdbaDuplicateIdError

data/lib/odba/version.rb

@@ -1,5 +1,5 @@
 #!/usr/bin/env ruby
 
 class Odba
-  VERSION = '1.2.2'
+  VERSION = '1.2.3'
 end

data/test/test_cache.rb

@@ -57,6 +57,36 @@ module ODBA
       @cache.instance_variable_set(:@peers, [peer])
       assert_equal(100, @cache.next_id)
     end
+
+    # And "cannot reach" is not only DRbError. A reference into a peer that
+    # has restarted or expired raises RangeError("invalid reference") from
+    # DRbObjectSpace#to_obj - a StandardError, not a DRbError. 1.2.2 named
+    # DRbError alone and so let it out: on 01.09.2026 three index rebuilds in
+    # oddb.org died on it, which 1.1.9's classless rescue had absorbed.
+    def test_a_stale_reference_into_a_peer_does_not_stop_the_allocation
+      @storage.should_receive(:next_id).and_return(100)
+      peer = flexmock("peer")
+      peer.should_receive(:reserve_next_id)
+        .and_raise(RangeError, "invalid reference")
+      @cache.instance_variable_set(:@peers, [peer])
+      assert_equal(100, @cache.next_id)
+    end
+
+    # The catch-all must not grow so wide that it takes the conflict with it
+    # again - the whole point of 1.2.2. One test for each direction.
+    def test_a_conflict_is_not_swallowed_by_the_catch_all
+      @storage.should_receive(:next_id).and_return(100, 101)
+      calls = 0
+      peer = flexmock("peer")
+      peer.should_receive(:reserve_next_id).and_return {
+        calls += 1
+        raise ODBA::OdbaDuplicateIdError, "taken" if calls == 1
+        true
+      }
+      @cache.instance_variable_set(:@peers, [peer])
+      assert_equal(101, @cache.next_id)
+      assert_equal(2, calls)
+    end
   end
 
   class TestCache < Test::Unit::TestCase

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: odba
 version: !ruby/object:Gem::Version
-  version: 1.2.2
+  version: 1.2.3
 platform: ruby
 authors:
 - Masaomi Hatakeyama, Zeno R.R. Davatz