diff --git a/CHANGELOG.md b/CHANGELOG.md
index 65159a69..6347293e 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,8 @@
 
 ### New Features
 
+* [#1036](https://github.com/toptal/chewy/issues/1036): Add `Chewy.close_client` and `Chewy::ElasticClient#close` to explicitly close connections to Elasticsearch, avoiding file descriptor leaks in long-lived multi-threaded processes (e.g. Sidekiq). ([@AlfonsoUceda][])
+
 ### Bug Fixes
 
 ### Changes
@@ -888,6 +890,7 @@
 [@AgeevAndrew]: https://github.com/AgeevAndrew
 [@aglushkov]: https://github.com/aglushkov
 [@AlexVPopov]: https://github.com/AlexVPopov
+[@AlfonsoUceda]: https://github.com/AlfonsoUceda
 [@AndreySavelyev]: https://github.com/AndreySavelyev
 [@afg419]: https://github.com/afg419
 [@arion]: https://github.com/arion
diff --git a/README.md b/README.md
index ae89d1eb..c3f4e938 100644
--- a/README.md
+++ b/README.md
@@ -133,6 +133,23 @@ development:
       ca_file: './tmp/http_ca.crt'
 ```
 
+### Closing connections
+
+`Chewy.client` is memoized per thread, so every thread that touches Chewy gets
+its own client with its own connections to Elasticsearch. Neither
+`elasticsearch-ruby` nor `elastic-transport` expose a public way to close those
+connections, so they are only released when Ruby's garbage collector reclaims
+the client. In long-lived, multi-threaded processes that frequently spawn and
+discard threads (e.g. Sidekiq, which replaces a thread on job failure), this can
+leak file descriptors.
+
+Use `Chewy.close_client` to close the current thread's connections and drop its
+client. The next `Chewy.client` call rebuilds a fresh one:
+
+```ruby
+Chewy.close_client
+```
+
 ### Index
 
 Create `app/chewy/users_index.rb` with User Index:
diff --git a/lib/chewy.rb b/lib/chewy.rb
index d573b654..1fad35b9 100644
--- a/lib/chewy.rb
+++ b/lib/chewy.rb
@@ -102,6 +102,21 @@ def client
       Chewy.current[:chewy_client] ||= Chewy::ElasticClient.new
     end
 
+    # Closes the current thread's client connections to Elasticsearch and
+    # drops the thread-local client, so the next `Chewy.client` call builds a
+    # fresh one.
+    #
+    # Useful in long-lived multi-threaded processes (e.g. Sidekiq) where the
+    # per-thread client would otherwise keep its connections open until the
+    # dead thread is garbage collected, leaking file descriptors.
+    def close_client
+      client = Chewy.current[:chewy_client]
+      return unless client
+
+      client.close
+      Chewy.current[:chewy_client] = nil
+    end
+
     # Sends wait_for_status request to ElasticSearch with status
     # defined in configuration.
     #
diff --git a/lib/chewy/elastic_client.rb b/lib/chewy/elastic_client.rb
index 41a985cc..15ec76fe 100644
--- a/lib/chewy/elastic_client.rb
+++ b/lib/chewy/elastic_client.rb
@@ -12,6 +12,21 @@ def initialize(elastic_client = self.class.build_es_client)
       @elastic_client = elastic_client
     end
 
+    # Closes the underlying connections to Elasticsearch.
+    #
+    # Neither elasticsearch-ruby nor elastic-transport expose a public method
+    # to close connections, so they are only released when Ruby's garbage
+    # collector reclaims the client instance. This reaches down to the Faraday
+    # connection of every transport connection and closes it explicitly, which
+    # is useful to avoid file descriptor leaks in long-lived processes that
+    # build a client per thread (e.g. Sidekiq workers).
+    def close
+      @elastic_client.transport.connections.each do |connection|
+        faraday = connection.connection
+        faraday.close if faraday.respond_to?(:close)
+      end
+    end
+
   private
 
     def method_missing(name, *args, **kwargs, &block)
diff --git a/spec/chewy/elastic_client_spec.rb b/spec/chewy/elastic_client_spec.rb
index 58cdc0cc..0c79814a 100644
--- a/spec/chewy/elastic_client_spec.rb
+++ b/spec/chewy/elastic_client_spec.rb
@@ -23,4 +23,24 @@
       Chewy.client.search({index: ['products'], body: {size: 0}}).to_a
     end
   end
+
+  describe '#close' do
+    let(:faraday_connection) { double(:faraday_connection) }
+    let(:connection) { double(:connection, connection: faraday_connection) }
+    let(:transport) { double(:transport, connections: [connection]) }
+    let(:elastic_client) { double(:elastic_client, transport: transport) }
+    let(:client) { described_class.new(elastic_client) }
+
+    it 'closes every underlying Faraday connection' do
+      allow(faraday_connection).to receive(:respond_to?).with(:close).and_return(true)
+      expect(faraday_connection).to receive(:close)
+      client.close
+    end
+
+    it 'skips connections that do not support close' do
+      allow(faraday_connection).to receive(:respond_to?).with(:close).and_return(false)
+      expect(faraday_connection).not_to receive(:close)
+      expect { client.close }.not_to raise_error
+    end
+  end
 end
diff --git a/spec/chewy_spec.rb b/spec/chewy_spec.rb
index 17e0511c..4a430e1a 100644
--- a/spec/chewy_spec.rb
+++ b/spec/chewy_spec.rb
@@ -58,6 +58,33 @@
     after { Chewy.current[:chewy_client] = initial_client }
   end
 
+  describe '.close_client' do
+    let!(:initial_client) { Chewy.current[:chewy_client] }
+
+    after { Chewy.current[:chewy_client] = initial_client }
+
+    context 'when a client exists for the current thread' do
+      let(:client) { instance_double(Chewy::ElasticClient) }
+
+      before { Chewy.current[:chewy_client] = client }
+
+      specify 'closes the client and clears the thread-local' do
+        expect(client).to receive(:close)
+        Chewy.close_client
+        expect(Chewy.current[:chewy_client]).to be_nil
+      end
+    end
+
+    context 'when no client exists for the current thread' do
+      before { Chewy.current[:chewy_client] = nil }
+
+      specify 'does nothing' do
+        expect { Chewy.close_client }.not_to raise_error
+        expect(Chewy.current[:chewy_client]).to be_nil
+      end
+    end
+  end
+
   describe '.create_indices' do
     before do
       stub_index(:cities)