I ran a long-running test that was navigating to a restricted set of web pages for a whole day. The browser process was kept alive the whole time. Memory dumps using memory-infra are taken at the beginning and at the end of the experiment.

There are about 21582 objects that are potential leaks or potential incorrect cache management with the following stack-frame:

      "operator new",
      "google::protobuf::internal::ArenaStringPtr::CreateInstanceNoArena",
      "google::protobuf::internal::ArenaStringPtr::AssignWithDefault",
      "metrics::TranslateEventProto::MergeFrom",
      "metrics::TranslateEventProto::TranslateEventProto",
      "std::vector<...>::_Reallocate",
      "std::vector<...>::_Reserve",
      "std::vector<...>::push_back",
      "translate::TranslateRankerImpl::AddTranslateEvent",
      "translate::TranslateManager::InitiateTranslation",
      "translate::ContentTranslateDriver::RegisterPage",
      "translate::mojom::ContentTranslateDriverStubDispatch::Accept",
      "mojo::InterfaceEndpointClient::HandleValidatedMessage",
      "mojo::FilterChain::Accept",
      "mojo::internal::MultiplexRouter::ProcessIncomingMessage",
      "mojo::internal::MultiplexRouter::Accept",
      "mojo::FilterChain::Accept",
      "mojo::Connector::ReadSingleMessage",
      "mojo::Connector::ReadAllAvailableMessages",
      "mojo::SimpleWatcher::OnHandleReady",
      "base::internal::InvokeHelper<1,void>::MakeItSo<void (__cdecl subresource_filter::SubframeNavigationFilteringThrottle::*const & __ptr64)(enum subresource_filter::SubframeNavigationFilteringThrottle::ThrottlingStage,enum subresource_filter::LoadPolicy) __ptr64,base::WeakPtr<subresource_filter::SubframeNavigationFilteringThrottle> const & __ptr64,enum subresource_filter::SubframeNavigationFilteringThrottle::ThrottlingStage const & __ptr64,enum subresource_filter::LoadPolicy>",
      "base::debug::TaskAnnotator::RunTask",
      "base::MessageLoop::RunTask",
      "base::MessageLoop::DoWork",
      "base::MessagePumpForUI::DoRunLoop",
      "base::MessagePumpWin::Run",
      "base::RunLoop::Run",
      "ChromeBrowserMainParts::MainMessageLoopRun",
      "content::BrowserMainLoop::RunMainMessageLoopParts",
      "content::BrowserMainRunnerImpl::Run",
      "content::BrowserMain",
      "content::RunNamedProcessTypeMain",
      "content::ContentMainRunnerImpl::Run",
      "service_manager::Main",
      "content::ContentMain",
      "ChromeMain",
      "MainDllLoader::Launch",
      "wWinMain",
      "__scrt_common_main_seh",
      "BaseThreadInitThunk",
      "RtlUserThreadStart",
      "[Thread: CrBrowserMain]"

The following line is adding the events:
  event_cache_.push_back(event);

  https://cs.chromium.org/chromium/src/components/translate/core/browser/translate_ranker_impl.cc?l=327&rcl=0f8c781859c2e2767a543d0238870c105a0ece88

They are cleared with:
  TranslateRankerImpl::FlushTranslateEvents(

which are called periodically.

Is is possible that these objects are never flushed under my experiment conditions?

The experiment was done with the following patch over telemetry:
  https://codereview.chromium.org/2906113002/