clickhouse-native 0.11.0 → 0.11.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 984a63d34de19388f777cfe71ecd0a844d22f0b67b89e9fddfe07c5c0ac1cc05
-  data.tar.gz: 78f829e40f715aef0264e1419cb9b06b97faab1ac38f33329c5da6ada1dd251b
+  metadata.gz: dbc7f982fc96317d6b6721bc7162e22d01eac4f4fc9edb307443d24fdee59032
+  data.tar.gz: 824150b24641414bfc0ccc4a6e37aa521b768bb73423c1bda9c1bfbefc466044
 SHA512:
-  metadata.gz: c579da78d2185108712c92e902217c2e508655533c2e78efc7e36ae358a826f809f41d884684b8fcb0e1b36749bc80be5f23b76c85ebf081423adb615d73928b
-  data.tar.gz: 93259646a6f150ed271202e3eaca075e0d1cf8d707aed5ac2ac9b530061de47cf1a65082289718857b367dd5b8d764c73cf270c34d080018ce8caac540d8bab6
+  metadata.gz: 6cda8a41d58816d63aa064ae3008b20980d35be6f695ff04433bc760ec28c0a286b707bdeb725d5691ac8245533c05d90df4c9c5b3eb13af6b860a179d9349db
+  data.tar.gz: caf00f391e65eb920a2d917316fa05c1569b7aac7c89f83a122b149c3305a3b0b40b6dc02afa5bd1cebcd78f0897700d5a2f9fbafb2c0d327da973dd3bc542f6

data/ext/clickhouse_native/client.cpp

@@ -872,6 +872,7 @@ struct ExecuteNoGVL {
     Client* client;
     const Query* query;
     std::exception_ptr err;
+    bool cancelled;
 };
 }  // namespace
 
@@ -885,11 +886,16 @@ static void* execute_no_gvl(void* data) {
     return nullptr;
 }
 
+// Unblock functions run on the *interrupting* thread while the blocked thread
+// is still inside Client::Execute(). CancelInFlight() only shuts the socket
+// down; ResetConnection() would free the very streams that thread is reading
+// (Thread#kill from Parallel.in_threads, Timeout, Sidekiq shutdown), which
+// reads back as a garbage packet type and then segfaults. The blocked call
+// returns EOF, raises, and the pool discards the client.
 static void execute_unblock(void* data) {
-    // The only safe abort clickhouse-cpp exposes is tearing the connection.
-    // On interrupt we kill the socket; the pool will discard this client.
     auto* a = static_cast<ExecuteNoGVL*>(data);
-    try { a->client->ResetConnection(); } catch (...) {}
+    a->cancelled = true;
+    try { a->client->CancelInFlight(); } catch (...) {}
 }
 
 static VALUE ch_client_execute(int argc, VALUE* argv, VALUE self) {
@@ -903,13 +909,20 @@ static VALUE ch_client_execute(int argc, VALUE* argv, VALUE self) {
     apply_default_settings(q, c);
     apply_settings(q, kwargs);
 
-    ExecuteNoGVL args{c->client.get(), &q, nullptr};
+    ExecuteNoGVL args{c->client.get(), &q, nullptr, false};
     rb_thread_call_without_gvl(execute_no_gvl, &args, execute_unblock, &args);
     if (args.err) {
         // clickhouse-cpp may leave the read stream partially consumed when the
         // server exception or an unsupported-type error is thrown mid-block.
         // Reset so the next call on this Client starts from a clean protocol.
-        try { c->client->ResetConnection(); } catch (...) {}
+        // Not after a cancel: that socket is already shut down and the caller
+        // discards the client, so the connect+handshake buys nothing. A
+        // Thread#kill or Timeout never gets here at all — step 5 of
+        // rb_thread_call_without_gvl delivers the interrupt before it returns —
+        // so this covers what is left: an unblock function that fires without a
+        // longjmp behind it, from a trap handler that does not raise or an
+        // interrupt Thread.handle_interrupt has deferred.
+        if (!args.cancelled) { try { c->client->ResetConnection(); } catch (...) {} }
         try { std::rethrow_exception(args.err); }
         catch (const std::exception& e) { raise_mapped_ex(e); }
     }
@@ -1006,6 +1019,7 @@ struct InsertNoGVL {
     const Block* block;
     const std::vector<std::pair<std::string, std::string>>* settings;
     std::exception_ptr err;
+    bool cancelled;
 };
 }  // namespace
 
@@ -1021,7 +1035,8 @@ static void* insert_no_gvl(void* data) {
 
 static void insert_unblock(void* data) {
     auto* a = static_cast<InsertNoGVL*>(data);
-    try { a->client->ResetConnection(); } catch (...) {}
+    a->cancelled = true;
+    try { a->client->CancelInFlight(); } catch (...) {}
 }
 
 static VALUE ch_client_insert_block(VALUE self, VALUE rb_table, VALUE rb_columns, VALUE rb_rows) {
@@ -1078,10 +1093,10 @@ static VALUE ch_client_insert_block(VALUE self, VALUE rb_table, VALUE rb_columns
             block.AppendColumn(names[i], cols[i]);
         }
 
-        InsertNoGVL args{c->client.get(), &table, &block, &c->default_settings, nullptr};
+        InsertNoGVL args{c->client.get(), &table, &block, &c->default_settings, nullptr, false};
         rb_thread_call_without_gvl(insert_no_gvl, &args, insert_unblock, &args);
         if (args.err) {
-            try { c->client->ResetConnection(); } catch (...) {}
+            if (!args.cancelled) { try { c->client->ResetConnection(); } catch (...) {} }
             try { std::rethrow_exception(args.err); }
             catch (const std::exception& e) { raise_mapped_ex(e); }
         }
@@ -1114,6 +1129,7 @@ struct QueryEachNoGVL {
     const Query* query;
     QueryEachState* state;
     std::exception_ptr err;
+    bool cancelled;
 };
 }  // namespace
 
@@ -1166,7 +1182,8 @@ static void* query_each_no_gvl(void* data) {
 static void query_each_unblock(void* data) {
     auto* a = static_cast<QueryEachNoGVL*>(data);
     a->state->aborted = true;
-    try { a->client->ResetConnection(); } catch (...) {}
+    a->cancelled = true;
+    try { a->client->CancelInFlight(); } catch (...) {}
 }
 
 static VALUE ch_client_query_each(int argc, VALUE* argv, VALUE self) {
@@ -1187,12 +1204,12 @@ static VALUE ch_client_query_each(int argc, VALUE* argv, VALUE self) {
         rb_thread_call_with_gvl(with_gvl_yield, &ya);
         return !state.aborted;
     });
-    QueryEachNoGVL args{c->client.get(), &q, &state, nullptr};
+    QueryEachNoGVL args{c->client.get(), &q, &state, nullptr, false};
 
     rb_thread_call_without_gvl(query_each_no_gvl, &args, query_each_unblock, &args);
 
     if (args.err) {
-        try { c->client->ResetConnection(); } catch (...) {}
+        if (!args.cancelled) { try { c->client->ResetConnection(); } catch (...) {} }
         if (state.exc_tag) rb_jump_tag(state.exc_tag);
         try { std::rethrow_exception(args.err); }
         catch (const std::exception& e) { raise_mapped_ex(e); }

data/ext/clickhouse_native/patches/0003-cancelable-socket.patch

@@ -0,0 +1,136 @@
+diff --git a/clickhouse/base/socket.cpp b/clickhouse/base/socket.cpp
+index 3bb1aa5..f188d3f 100644
+--- a/clickhouse/base/socket.cpp
++++ b/clickhouse/base/socket.cpp
+@@ -345,6 +345,19 @@ Socket::~Socket() {
+     Close();
+ }
+ 
++void Socket::Cancel() {
++    if (handle_ == INVALID_SOCKET) {
++        return;
++    }
++    // shutdown(), not close(): the descriptor stays valid, so a concurrent
++    // reader returns EOF instead of racing a reused fd number.
++#if defined(_win_)
++    shutdown(handle_, SD_BOTH);
++#else
++    ::shutdown(handle_, SHUT_RDWR);
++#endif
++}
++
+ void Socket::Close() {
+     CloseSocket(handle_);
+     handle_ = INVALID_SOCKET;
+diff --git a/clickhouse/base/socket.h b/clickhouse/base/socket.h
+index 9bd9ca3..0234b1e 100644
+--- a/clickhouse/base/socket.h
++++ b/clickhouse/base/socket.h
+@@ -80,6 +80,12 @@ public:
+ 
+     virtual std::unique_ptr<InputStream> makeInputStream() const = 0;
+     virtual std::unique_ptr<OutputStream> makeOutputStream() const = 0;
++
++    /// Unblock a thread parked in recv()/send() on this socket. Unlike
++    /// closing or replacing the socket it destroys nothing, so it is safe
++    /// to call from another thread while the owner is mid-read: the owner
++    /// sees EOF and unwinds through its own frames.
++    virtual void Cancel() {}
+ };
+ 
+ 
+@@ -123,6 +129,8 @@ public:
+     std::unique_ptr<InputStream> makeInputStream() const override;
+     std::unique_ptr<OutputStream> makeOutputStream() const override;
+ 
++    void Cancel() override;
++
+ protected:
+     Socket(const Socket&) = delete;
+     Socket& operator = (const Socket&) = delete;
+diff --git a/clickhouse/client.cpp b/clickhouse/client.cpp
+index c9732f4..88c9b84 100644
+--- a/clickhouse/client.cpp
++++ b/clickhouse/client.cpp
+@@ -9,6 +9,7 @@
+ #include "columns/factory.h"
+ 
+ #include <cassert>
++#include <mutex>
+ #include <optional>
+ #include <sstream>
+ #include <system_error>
+@@ -226,6 +227,8 @@ public:
+ 
+     void ResetConnection();
+ 
++    void CancelInFlight();
++
+     void ResetConnectionEndpoint();
+ 
+     const ServerInfo& GetServerInfo() const;
+@@ -313,6 +316,11 @@ private:
+ 
+     std::unique_ptr<SocketFactory> socket_factory_;
+ 
++    /// Guards socket_ against CancelInFlight() on another thread. Covers the
++    /// swap in ResetConnection() and the displaced socket's destructor, so a
++    /// canceller can never shutdown() an fd that is closing or already reissued.
++    std::mutex socket_mutex_;
++
+     std::unique_ptr<InputStream> input_;
+     std::unique_ptr<OutputStream> output_;
+     std::unique_ptr<SocketBase> socket_;
+@@ -610,8 +618,24 @@ void Client::Impl::Ping() {
+     }
+ }
+ 
++void Client::Impl::CancelInFlight() {
++    std::lock_guard<std::mutex> lock(socket_mutex_);
++    if (socket_) {
++        socket_->Cancel();
++    }
++}
++
+ void Client::Impl::ResetConnection() {
+-    InitializeStreams(socket_factory_->connect(options_, current_endpoint_.value()));
++    auto fresh = socket_factory_->connect(options_, current_endpoint_.value());
++    {
++        // InitializeStreams swaps, so `fresh` comes back holding the *old*
++        // socket; reset it here so its destructor runs under the lock too.
++        // Left to itself it would die in this function's full-expression,
++        // outside any lock, which is the window CancelInFlight() must not hit.
++        std::lock_guard<std::mutex> lock(socket_mutex_);
++        InitializeStreams(std::move(fresh));
++        fresh.reset();
++    }
+     state_ = State::Idle;
+ 
+     if (!Handshake()) {
+@@ -1408,6 +1432,10 @@ void Client::ResetConnection() {
+     impl_->ResetConnection();
+ }
+ 
++void Client::CancelInFlight() {
++    impl_->CancelInFlight();
++}
++
+ void Client::ResetConnectionEndpoint() {
+     impl_->ResetConnectionEndpoint();
+ }
+diff --git a/clickhouse/client.h b/clickhouse/client.h
+index 6c6ff53..27be58c 100644
+--- a/clickhouse/client.h
++++ b/clickhouse/client.h
+@@ -326,6 +326,11 @@ public:
+     /// Reset connection with initial params.
+     void ResetConnection();
+ 
++    /// Unblock a thread parked in an Execute()/Insert() socket read.
++    /// Destroys nothing, so unlike ResetConnection() it is safe to call
++    /// from another thread while a query is still in flight.
++    void CancelInFlight();
++
+     const ServerInfo& GetServerInfo() const;
+ 
+     /// Get current connected endpoint.

data/lib/clickhouse_native/pool.rb

@@ -22,8 +22,8 @@ module ClickhouseNative
       end
     end
 
-    # On exception, discard the client rather than reuse it: an error
-    # path leaves the socket in an unknown state. The C++ binding issues
+    # On an aborted operation, discard the client rather than reuse it: the
+    # socket is left in an unknown state. The C++ binding issues
     # ResetConnection, but a subsequent send can still surface buffered
     # protocol errors from the prior aborted operation — those get
     # attributed to whatever SQL we tried next, producing misleading log
@@ -44,15 +44,18 @@ module ClickhouseNative
     # before running the query, so most stale connections never surface as
     # a ConnectionError here at all. This retry still covers the residual
     # race (socket dies between the ping and the query).
+    #
+    # A bare `rescue` is not enough to spot an abandoned query. Timeout and
+    # Sidekiq shutdown raise off Exception rather than StandardError, and
+    # Thread#kill raises nothing at all — Parallel.in_threads kills every
+    # sibling worker the moment one of them fails. Miss those and the
+    # connection goes back in with the server still streaming a response at
+    # it; the next checkout reads that leftover as its own, which surfaces
+    # as a bogus packet type far from here.
     def with
       attempts = 0
       begin
-        @pool.with do |client|
-          yield client
-        rescue
-          @pool.discard_current_connection(&:close)
-          raise
-        end
+        @pool.with { |client| discard_unless_clean { yield client } }
       rescue ConnectionError
         attempts += 1
         retry if attempts == 1
@@ -91,5 +94,30 @@ module ClickhouseNative
     def describe_table(table, db_name: nil)
       with { |c| c.describe_table(table, db_name:) }
     end
+
+    private
+
+    # Keep the client only when the block either finished or left of its own
+    # accord. A `break` out of a streaming read is deliberate, and the binding
+    # has already reconnected the client on its way out (query_each resets the
+    # connection before rb_jump_tag), so it is good to reuse — discarding would
+    # buy a second connect on top of the one already paid. An exception is an
+    # abort, and so is Thread#kill: it raises nothing, so no rescue ever sees
+    # it, and its only trace is an "aborting" thread while ensure runs.
+    def discard_unless_clean
+      finished = false
+      begin
+        result = yield
+        finished = true
+        result
+      rescue Exception # rubocop:disable Lint/RescueException
+        @pool.discard_current_connection(&:close)
+        raise
+      ensure
+        if !finished && Thread.current.status == "aborting"
+          @pool.discard_current_connection(&:close)
+        end
+      end
+    end
   end
 end

data/lib/clickhouse_native/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module ClickhouseNative
-  VERSION = "0.11.0"
+  VERSION = "0.11.1"
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: clickhouse-native
 version: !ruby/object:Gem::Version
-  version: 0.11.0
+  version: 0.11.1
 platform: ruby
 authors:
 - Yuri Smirnov
@@ -36,6 +36,7 @@ files:
 - ext/clickhouse_native/extconf.rb
 - ext/clickhouse_native/patches/0001-preserve-declared-column-type.patch
 - ext/clickhouse_native/patches/0002-carry-settings-into-insert.patch
+- ext/clickhouse_native/patches/0003-cancelable-socket.patch
 - ext/clickhouse_native/vendor/clickhouse-cpp/.clang-format
 - ext/clickhouse_native/vendor/clickhouse-cpp/.git
 - ext/clickhouse_native/vendor/clickhouse-cpp/.gitattributes