[k2] rpc client API by T-y-c-o-o-n · Pull Request #1635 · VKCOM/kphp

T-y-c-o-o-n · 2026-06-15T10:37:24Z

No description provided.

…GAIN` errno if response is not ready

…e std::string_view instead of string, removed m_id from query_handle

…nse()

…ator

…sponse_awaiter_tasks

…:STATIC_BUFFER_LENGTH`; added `if collect_responses_extra_info` check

apolyakov · 2026-07-27T06:48:29Z

+  return {descriptor};
+}
+
+inline std::expected<size_t, int32_t> rpc_get_response_size(uint64_t rpc_d) noexcept {


Let's use k2::descriptor instead of uint64_t

apolyakov · 2026-07-27T06:53:22Z

 int32_t k2_component_access(size_t name_len, const char* name);

+/**
+ * Try to send rpc request. In case of success write descriptor of rpc request to `rpc_d`, otherwise return `errno` != 0,


Could you please make these docs more clear?

For example, for k2_rpc_send_request it might look like:

* Try to send an RPC request. On success, write the request descriptor to `rpc_d`; * this descriptor is later used when calling `k2_rpc_fetch_response`. * Otherwise, return a non-zero `errno`. ...

Please, take a look at other API functions you've added and check their docs too

apolyakov · 2026-07-27T06:57:59Z

  int64_t current_query_id{kphp::rpc::VALID_QUERY_ID_RANGE_START};

-  kphp::stl::unordered_map<int64_t, kphp::coro::shared_task<std::optional<string>>, kphp::memory::script_allocator> response_awaiter_tasks;
+  kphp::stl::unordered_map<int64_t, kphp::coro::shared_task<std::expected<kphp::stl::vector<std::byte, kphp::memory::script_allocator>, int32_t>>,


This is not a good idea to have shared_task returning something like vector. shared_task returns a copy of its result on each co_await, so you'll end up with multiple copies of the vector

apolyakov · 2026-07-27T07:00:52Z

+        m_deadline{deadline} {}
+
+  template<std::invocable<size_t> B>
+  requires std::is_convertible_v<std::invoke_result_t<B, size_t>, std::span<std::byte>>


Let's make it std::same_as<std::span<byte>>

apolyakov · 2026-07-27T07:07:18Z

+
+  template<std::invocable<size_t> B>
+  requires std::is_convertible_v<std::invoke_result_t<B, size_t>, std::span<std::byte>>
+  auto get_ready_response(B response_buffer_provider) noexcept -> std::expected<std::span<std::byte>, int32_t>;


Why do we need to accept this parameter by value?

apolyakov · 2026-07-27T07:11:15Z

+
+  template<std::invocable<size_t> B>
+  requires std::is_convertible_v<std::invoke_result_t<B, size_t>, std::span<std::byte>>
+  auto get_response(B response_buffer_provider) noexcept -> kphp::coro::task<std::expected<std::span<std::byte>, int32_t>>;


Just a note to think about: we can get rid of this method. Instead, we might have an interface like this:

kphp::rpc::query q{create_some_query()}; co_await q(my_response_buffer_provider);

What do you think?

apolyakov · 2026-07-27T07:14:57Z

+  }
+  size_t response_size{*response_size_exp};
+
+  std::span<std::byte> response_buffer{std::invoke(response_buffer_provider, response_size)};


I think we need to check response_buffer's size here

apolyakov · 2026-07-27T07:16:16Z

+requires std::is_convertible_v<std::invoke_result_t<B, size_t>, std::span<std::byte>>
+auto query::get_response(B response_buffer_provider) noexcept -> kphp::coro::task<std::expected<std::span<std::byte>, int32_t>> {
+  if (m_descriptor == k2::INVALID_PLATFORM_DESCRIPTOR) {
+    co_return std::unexpected{TL_ERROR_INTERNAL};


We can use TL_ERROR_INVALID_CONNECTION_ID to improve observability

apolyakov · 2026-07-27T07:17:55Z

+  kphp::coro::io_scheduler& m_scheduler{kphp::coro::io_scheduler::get()};
+  std::chrono::nanoseconds timeout{kphp::time::remaining(m_deadline)};
+
+  // TODO DISCUSS: may be completely remove timeout monitoring in kphp and leave it in k2-node's rpc client ???


Seems reasonable now. There are two ways the platform can notify us about RPC timeout:

close the stream -> aren't there other errors that can lead to stream being closed?

write some "there was a timeout" answer to the stream -> requires a bit more complex processing since such an answer will vary depending on RPC kind we use

apolyakov · 2026-07-27T07:25:20Z

+/**
+ * Calculate time remaining to the deadline.
+ */
+inline std::chrono::nanoseconds remaining(std::chrono::steady_clock::time_point deadline) noexcept {


What do you think about following implementation:

/** * Returns the current steady-clock time point from the K2 platform. */ inline std::chrono::steady_clock::time_point now() noexcept /* iff k2::instant is noexcept */ { k2::TimePoint now_instant{}; k2::instant(std::addressof(now_instant)); return std::chrono::steady_clock::time_point{std::chrono::nanoseconds{now_instant.time_point_ns}}; } /** * Calculate time remaining until the deadline. */ template <typename Duration> inline auto remaining(std::chrono::time_point<std::chrono::steady_clock, Duration> deadline) noexcept { return deadline - now(); } /** * Converts a timeout to the time point at which it elapses, i.e. the deadline. */ template <typename Rep, typename Period> inline auto expires_at(std::chrono::duration<Rep, Period> timeout) noexcept { return now() + timeout; }

It makes client code easier by removing mandatory conversions from any duration to nanoseconds

apolyakov · 2026-07-31T09:58:05Z

-inline kphp::coro::task<array<int64_t>> f$rpc_send_requests(string actor, array<mixed> tl_objects, Optional<double> timeout, bool ignore_answer,
-                                                            class_instance<C$KphpRpcRequestsExtraInfo> requests_extra_info,
-                                                            bool need_responses_extra_info) noexcept {
+inline array<int64_t> f$rpc_send_requests(string actor, array<mixed> tl_objects, Optional<double> timeout, bool ignore_answer,


As it's not longer a coroutine, we don't need to pass parameters by value

apolyakov · 2026-07-31T10:00:07Z

+        m_deadline{deadline} {}
+
+  template<std::invocable<size_t> B>
+  requires std::is_same_v<std::invoke_result_t<B, size_t>, std::span<std::byte>>


missing includes for concepts and type_traits

apolyakov · 2026-07-31T10:01:21Z

+
+inline auto query::send(std::string_view actor, std::chrono::milliseconds timeout,
+                        std::span<const std::byte> request_buffer) noexcept -> std::expected<query, int32_t> {
+  auto descriptor_exp{k2::rpc_send_request(actor, request_buffer, RpcKind::TL_RPC)};


I think send needs to accept RPC kind as a parameter

apolyakov · 2026-07-31T10:03:46Z

+    co_return std::unexpected{TL_ERROR_INVALID_CONNECTION_ID};
+  }
+
+  kphp::coro::io_scheduler& m_scheduler{kphp::coro::io_scheduler::get()};


nit: no need to get the scheduler unless it's really needed

apolyakov · 2026-07-31T10:07:34Z

+  auto timeout{kphp::time::remaining(m_deadline)};
+
+  if (timeout <= std::chrono::nanoseconds::zero()) {
+    co_return std::move(*this).get_ready_response<B>(std::move(response_buffer_provider));


nit: I think it's better to declare get_ready_response and static constexpr lambda inside response

apolyakov · 2026-07-31T10:18:20Z

-        std::swap(tl_storer, rpc_server_instance_st.tl_storer);
-      }
-    }};
+  // if we have to allocate memory for request buffer, it will be in this vector, and it will be freed at the end of the function


I think it's better to use the tl_storer buffer for this

apolyakov · 2026-07-31T10:20:01Z

-    auto fetch_task{kphp::component::fetch_response(stream, kphp::component::read_ext::append(*opt_response))};
-    if (auto expected{co_await kphp::coro::io_scheduler::get().schedule(std::move(fetch_task), timeout)}; !expected) [[unlikely]] {
-      opt_response = std::nullopt;
+  static constexpr auto awaiter_coroutine{[](query q, int64_t query_id, bool collect_responses_extra_info) mutable noexcept


Why is it mutable?

apolyakov · 2026-07-31T10:21:27Z

+                                          -> kphp::coro::shared_task<std::expected<string, int32_t>> {
+    string resp_buf{};
+    const auto response_buffer_provider{[&resp_buf](size_t size) noexcept -> std::span<std::byte> {
+      resp_buf.assign(size, true);


reserve_at_least better expresses what you do here

apolyakov · 2026-07-31T10:22:07Z

+      return {reinterpret_cast<std::byte*>(resp_buf.buffer()), size};
+    }};
+
+    auto fetch_task{std::move(q).response<decltype(response_buffer_provider)>(response_buffer_provider)};


No need for explicit decltype here

apolyakov · 2026-07-31T10:22:35Z

-        auto fetch_task{kphp::component::fetch_response(stream, [](std::span<const std::byte>) noexcept {})};
-        std::ignore = co_await kphp::coro::io_scheduler::get().schedule(std::move(fetch_task), timeout);
-      }};
+  static constexpr auto ignore_answer_awaiter_coroutine{[](query q) noexcept -> kphp::coro::shared_task<> {


Do we even need it now?

T-y-c-o-o-n force-pushed the n.siniachenko/k2/rpc-client-boost branch from 156e9ec to 222adae Compare June 15, 2026 11:57

T-y-c-o-o-n commented Jun 17, 2026

View reviewed changes

Comment thread runtime-light/stdlib/rpc/rpc-api.cpp

Comment thread runtime-light/stdlib/rpc/rpc-queue-functions.h Outdated

apolyakov requested changes Jun 18, 2026

View reviewed changes

T-y-c-o-o-n force-pushed the n.siniachenko/k2/rpc-client-boost branch from a342aa9 to e77636c Compare June 21, 2026 19:19

apolyakov reviewed Jun 29, 2026

View reviewed changes

Comment thread runtime-light/k2-platform/k2-header.h Outdated

Comment thread runtime-light/k2-platform/k2-api.h Outdated

Comment thread runtime-light/stdlib/time/util.h Outdated

T-y-c-o-o-n force-pushed the n.siniachenko/k2/rpc-client-boost branch 2 times, most recently from 0c1f4cc to 44d348e Compare July 17, 2026 13:14

apolyakov requested changes Jul 21, 2026

View reviewed changes

T-y-c-o-o-n changed the title ~~[k2] rpc client boost~~ [k2] rpc client API Jul 22, 2026

apolyakov reviewed Jul 22, 2026

View reviewed changes

Comment thread runtime-light/stdlib/rpc/rpc-query-handle.h Outdated

Comment thread runtime-light/stdlib/rpc/rpc-query-handle.h Outdated

Nikita Siniachenko added 20 commits July 24, 2026 18:47

first impl

bf39efd

added kphp rpc client bench

9daec6b

break in switch + different debug exit codes

cbdbae8

null timeout in bench

b68b51b

rpc send functions now common functions instead of coroutines

7cbcb7c

remove one coroutine on rpc response fetch

c609910

removed test trash

19e0065

error handling

1d38897

rpc_queue_push

4f8b439

kphp::rpc::request_info

31555bf

fmt

2340e15

added RpcKind + better docs

dbc67fd

query_handle

d5353ae

send_and_get_handle

0f1c8e2

refactored query_handle a bit

896ba9c

refactorrr

496db0f

k2_rpc_get_response_size() and k2_rpc_fetch_response() now return `EA…

f219b9e

…GAIN` errno if response is not ready

[EXPERIMENTAL] removed ignore_answer_coroutine

68e2711

regularize_extra_headers

1dad070

added query_handle::wait_for_response

0ca689d

Nikita Siniachenko added 20 commits July 24, 2026 18:48

fixed zeroing query_handle in rpc_queue_push(..)

5d0f66b

insert ignore_answer requests to rpc_query_handles

44c20e3

issues

fd9c00a

fmt

cc232f0

fmt

5c6e40d

refactored kphp::rpc::query_handle

e21663b

removed rpc-query-handle.cpp

222c87b

fixed nullptr request_buffer, fixed <=0 timeout handling, error becam…

e812bfe

…e std::string_view instead of string, removed m_id from query_handle

query_handle: response type now std::span<std::byte>

f2a9369

time utils fixes

a70b830

fmt

4576a4a

fmt

99d7de1

query_handle -> query; removed error description for query::get_respo…

89b6bc6

…nse()

kphp::rpc::query::get_ready_response() takes templated response_alloc…

91ab571

…ator

TODO

0788026

ResponseAllocator and ResponseDeleter

f957b0b

ResponseBufferProvider; removed query::wait_for_response; returned re…

0adcd48

…sponse_awaiter_tasks

request_buffer is allocated for requests more than `StringLibContext:…

47f98cb

…:STATIC_BUFFER_LENGTH`; added `if collect_responses_extra_info` check

fmt

fcf22b8

fmt

6435238

T-y-c-o-o-n force-pushed the n.siniachenko/k2/rpc-client-boost branch from df743db to 6435238 Compare July 24, 2026 15:48

Nikita Siniachenko added 2 commits July 24, 2026 19:33

fmt

d0c69df

k2_rpc_fetch_response may return ENOBUFS

bb2047b

apolyakov requested changes Jul 27, 2026

View reviewed changes

Nikita Siniachenko added 3 commits July 29, 2026 00:03

fixed issues

e6ca343

better docs

f7936b0

removed extra includes and comments

1c02dca

apolyakov reviewed Jul 31, 2026

View reviewed changes

Uh oh!

Conversation

T-y-c-o-o-n commented Jun 15, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

apolyakov Jul 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

apolyakov Jul 27, 2026 •

edited

Loading