|
| 1 | +defmodule AlexClaw.Skills.CircuitBreaker do |
| 2 | + @moduledoc """ |
| 3 | + OTP circuit breaker per skill. Wraps skill execution transparently — |
| 4 | + skills have zero awareness of the breaker. |
| 5 | +
|
| 6 | + States: :closed (normal) → :open (failing) → :half_open (testing) |
| 7 | +
|
| 8 | + Uses ETS for lock-free reads on the hot path and Process.send_after |
| 9 | + for reset timers. One GenServer per skill, managed by CircuitBreakerSupervisor. |
| 10 | + """ |
| 11 | + use GenServer |
| 12 | + require Logger |
| 13 | + |
| 14 | + @ets_table :circuit_breakers |
| 15 | + @max_failures Application.compile_env(:alex_claw, [:circuit_breaker, :max_failures], 3) |
| 16 | + @reset_timeout Application.compile_env(:alex_claw, [:circuit_breaker, :reset_timeout], :timer.minutes(5)) |
| 17 | + |
| 18 | + # --- Client API --- |
| 19 | + |
| 20 | + @doc "Start a circuit breaker GenServer for a skill." |
| 21 | + @spec start_link(String.t()) :: GenServer.on_start() |
| 22 | + def start_link(skill_name) do |
| 23 | + GenServer.start_link(__MODULE__, skill_name, name: via(skill_name)) |
| 24 | + end |
| 25 | + |
| 26 | + @doc """ |
| 27 | + Transparent wrapper for skill execution. Checks circuit state, |
| 28 | + executes the function if allowed, records the result. |
| 29 | + Skills see no difference — the breaker is invisible. |
| 30 | + """ |
| 31 | + @spec call(String.t(), (-> {:ok, any()} | {:error, any()})) :: |
| 32 | + {:ok, any()} | {:error, any()} | {:error, :circuit_open} |
| 33 | + def call(skill_name, fun) do |
| 34 | + AlexClaw.Skills.CircuitBreakerSupervisor.ensure_started(skill_name) |
| 35 | + |
| 36 | + if allow?(skill_name) do |
| 37 | + case fun.() do |
| 38 | + {:ok, result} -> |
| 39 | + record_success(skill_name) |
| 40 | + {:ok, result} |
| 41 | + |
| 42 | + {:error, reason} -> |
| 43 | + record_failure(skill_name, reason) |
| 44 | + {:error, reason} |
| 45 | + end |
| 46 | + else |
| 47 | + {:error, :circuit_open} |
| 48 | + end |
| 49 | + end |
| 50 | + |
| 51 | + @doc "Check if a skill is allowed to execute. Pure ETS read, no GenServer call." |
| 52 | + @spec allow?(String.t()) :: boolean() |
| 53 | + def allow?(skill_name) do |
| 54 | + case :ets.lookup(@ets_table, skill_name) do |
| 55 | + [{_, :open, _, _, _}] -> false |
| 56 | + _ -> true |
| 57 | + end |
| 58 | + end |
| 59 | + |
| 60 | + @doc "Return the circuit state for a skill." |
| 61 | + @spec state(String.t()) :: {atom(), non_neg_integer(), any()} | :unknown |
| 62 | + def state(skill_name) do |
| 63 | + case :ets.lookup(@ets_table, skill_name) do |
| 64 | + [{_, state, count, last_error, _}] -> {state, count, last_error} |
| 65 | + [] -> :unknown |
| 66 | + end |
| 67 | + end |
| 68 | + |
| 69 | + @doc "Force-reset a circuit to :closed. Used when a dynamic skill is reloaded." |
| 70 | + @spec reset(String.t()) :: :ok |
| 71 | + def reset(skill_name) do |
| 72 | + case Registry.lookup(AlexClaw.CircuitBreakerRegistry, skill_name) do |
| 73 | + [{pid, _}] -> GenServer.call(pid, :reset) |
| 74 | + [] -> :ok |
| 75 | + end |
| 76 | + end |
| 77 | + |
| 78 | + # --- GenServer callbacks --- |
| 79 | + |
| 80 | + @impl true |
| 81 | + def init(skill_name) do |
| 82 | + :ets.insert(@ets_table, {skill_name, :closed, 0, nil, DateTime.utc_now()}) |
| 83 | + {:ok, %{skill_name: skill_name, timer_ref: nil}} |
| 84 | + end |
| 85 | + |
| 86 | + @impl true |
| 87 | + def handle_cast({:record_failure, reason}, state) do |
| 88 | + {_, current_state, count, _, _} = lookup!(state.skill_name) |
| 89 | + new_count = count + 1 |
| 90 | + |
| 91 | + case current_state do |
| 92 | + :closed when new_count >= @max_failures -> |
| 93 | + transition(state, :open, new_count, reason) |
| 94 | + |
| 95 | + :half_open -> |
| 96 | + transition(state, :open, new_count, reason) |
| 97 | + |
| 98 | + _ -> |
| 99 | + :ets.insert(@ets_table, {state.skill_name, current_state, new_count, reason, DateTime.utc_now()}) |
| 100 | + {:noreply, state} |
| 101 | + end |
| 102 | + end |
| 103 | + |
| 104 | + def handle_cast(:record_success, state) do |
| 105 | + {_, current_state, _, _, _} = lookup!(state.skill_name) |
| 106 | + |
| 107 | + case current_state do |
| 108 | + :half_open -> |
| 109 | + transition(state, :closed, 0, nil) |
| 110 | + |
| 111 | + _ -> |
| 112 | + :ets.insert(@ets_table, {state.skill_name, current_state, 0, nil, DateTime.utc_now()}) |
| 113 | + {:noreply, state} |
| 114 | + end |
| 115 | + end |
| 116 | + |
| 117 | + @impl true |
| 118 | + def handle_call(:reset, _from, state) do |
| 119 | + cancel_timer(state.timer_ref) |
| 120 | + :ets.insert(@ets_table, {state.skill_name, :closed, 0, nil, DateTime.utc_now()}) |
| 121 | + Logger.info("[CircuitBreaker] #{state.skill_name} manually reset to :closed") |
| 122 | + {:reply, :ok, %{state | timer_ref: nil}} |
| 123 | + end |
| 124 | + |
| 125 | + @impl true |
| 126 | + def handle_info(:half_open, state) do |
| 127 | + :ets.insert(@ets_table, {state.skill_name, :half_open, 0, nil, DateTime.utc_now()}) |
| 128 | + Logger.info("[CircuitBreaker] #{state.skill_name} → :half_open (testing)") |
| 129 | + {:noreply, %{state | timer_ref: nil}} |
| 130 | + end |
| 131 | + |
| 132 | + # --- Internal --- |
| 133 | + |
| 134 | + defp record_success(skill_name) do |
| 135 | + case Registry.lookup(AlexClaw.CircuitBreakerRegistry, skill_name) do |
| 136 | + [{pid, _}] -> GenServer.cast(pid, :record_success) |
| 137 | + [] -> :ok |
| 138 | + end |
| 139 | + end |
| 140 | + |
| 141 | + defp record_failure(skill_name, reason) do |
| 142 | + case Registry.lookup(AlexClaw.CircuitBreakerRegistry, skill_name) do |
| 143 | + [{pid, _}] -> GenServer.cast(pid, {:record_failure, reason}) |
| 144 | + [] -> :ok |
| 145 | + end |
| 146 | + end |
| 147 | + |
| 148 | + defp transition(state, :open, count, reason) do |
| 149 | + cancel_timer(state.timer_ref) |
| 150 | + :ets.insert(@ets_table, {state.skill_name, :open, count, reason, DateTime.utc_now()}) |
| 151 | + timer_ref = Process.send_after(self(), :half_open, @reset_timeout) |
| 152 | + |
| 153 | + Logger.error( |
| 154 | + "[CircuitBreaker] #{state.skill_name} → :open after #{count} failures. " <> |
| 155 | + "Last error: #{inspect(reason)}. Auto-retry in #{div(@reset_timeout, 60_000)} min." |
| 156 | + ) |
| 157 | + |
| 158 | + notify_opened(state.skill_name, count, reason) |
| 159 | + {:noreply, %{state | timer_ref: timer_ref}} |
| 160 | + end |
| 161 | + |
| 162 | + defp transition(state, :closed, count, _reason) do |
| 163 | + cancel_timer(state.timer_ref) |
| 164 | + :ets.insert(@ets_table, {state.skill_name, :closed, count, nil, DateTime.utc_now()}) |
| 165 | + |
| 166 | + Logger.info("[CircuitBreaker] #{state.skill_name} → :closed (recovered)") |
| 167 | + |
| 168 | + notify_closed(state.skill_name) |
| 169 | + {:noreply, %{state | timer_ref: nil}} |
| 170 | + end |
| 171 | + |
| 172 | + defp lookup!(skill_name) do |
| 173 | + case :ets.lookup(@ets_table, skill_name) do |
| 174 | + [entry] -> entry |
| 175 | + [] -> {skill_name, :closed, 0, nil, DateTime.utc_now()} |
| 176 | + end |
| 177 | + end |
| 178 | + |
| 179 | + defp cancel_timer(nil), do: :ok |
| 180 | + defp cancel_timer(ref), do: Process.cancel_timer(ref) |
| 181 | + |
| 182 | + defp via(skill_name) do |
| 183 | + {:via, Registry, {AlexClaw.CircuitBreakerRegistry, skill_name}} |
| 184 | + end |
| 185 | + |
| 186 | + defp notify_opened(skill_name, count, reason) do |
| 187 | + Task.Supervisor.start_child(AlexClaw.TaskSupervisor, fn -> |
| 188 | + AlexClaw.Gateway.send_message( |
| 189 | + "⚡ Circuit *OPEN* for skill `#{skill_name}`\n" <> |
| 190 | + "#{count} consecutive failures. Auto-retry in #{div(@reset_timeout, 60_000)} min.\n" <> |
| 191 | + "Last error: `#{String.slice(inspect(reason), 0, 200)}`" |
| 192 | + ) |
| 193 | + end) |
| 194 | + end |
| 195 | + |
| 196 | + defp notify_closed(skill_name) do |
| 197 | + Task.Supervisor.start_child(AlexClaw.TaskSupervisor, fn -> |
| 198 | + AlexClaw.Gateway.send_message("✅ Circuit *CLOSED* for skill `#{skill_name}` — recovered.") |
| 199 | + end) |
| 200 | + end |
| 201 | +end |
0 commit comments