test: coverage for codec/transport/otel/grpc/lease/gc/vm (#69)

Author: markovejnovic

coveralls.json

@@ -1,5 +1,6 @@
 {
   "skip_files": [
+    "lib/hyper/firecracker/api/codec.ex",
     "lib/hyper/firecracker/api/operations/",
     "lib/hyper/firecracker/api/schemas/",
     "lib/mix/"

lib/hyper/img/db/gc.ex

@@ -24,16 +24,18 @@ defmodule Hyper.Img.Db.Gc do
   **Publish contract:** write a layer's file to the medium before inserting its
   `blobs` row (ideally write-temp then atomic rename); the grace period is
   insurance, not a substitute.
+
+  The delete step itself lives in `Hyper.Img.Db.Gc.Prune`.
   """
 
   use GenServer
   use OpenTelemetryDecorator
   require Logger
-  import Ecto.Query
 
   alias Hyper.Cfg.Gc, as: Config
   alias Hyper.Cluster.Routing
-  alias Hyper.Img.Db.{Blob, ImageLayer, Repo}
+  alias Hyper.Img.Db.Blob
+  alias Hyper.Img.Db.Gc.Prune
   alias Hyper.Img.Db.Gc.Sweep
   alias Hyper.Node.Layer.Repo, as: LayerRepo
 
@@ -116,8 +118,6 @@ defmodule Hyper.Img.Db.Gc do
   # role, monitor noise, etc.) rather than crashing an in-flight sweep.
   def handle_info(_msg, state), do: {:noreply, state}
 
-  ## Internals
-
   @spec acquire(t()) :: t()
   defp acquire(state) do
     case Horde.Registry.register(Routing.name(), @singleton_key, nil) do
@@ -136,10 +136,13 @@ defmodule Hyper.Img.Db.Gc do
   @decorate with_span("Hyper.Img.Db.Gc.scan_one_batch")
   defp scan_one_batch(%__MODULE__{sweep: sweep} = state) do
     limit = state.config.batch_size
-    batch = with_low_priority(state, fn -> Blob.present_after(sweep.cursor, limit) end)
-    {sweep, missing} = Sweep.absorb(sweep, batch, &presence/1)
 
-    {pruned, pruned_bytes, dangling} = maybe_prune(state, missing)
+    batch =
+      Prune.with_low_priority(state.config, fn -> Blob.present_after(sweep.cursor, limit) end)
+
+    {sweep, missing} = Sweep.absorb(sweep, batch, &Prune.presence/1)
+
+    {pruned, pruned_bytes, dangling} = Prune.execute(state.config, missing)
     sweep = Sweep.record_prune(sweep, pruned, pruned_bytes, dangling)
 
     if Sweep.continue?(batch, limit) do
@@ -156,120 +159,4 @@ defmodule Hyper.Img.Db.Gc do
       %{state | sweep: nil, last_sweep: sweep}
     end
   end
-
-  # Re-check the medium is still mounted before acting on a page's "missing"
-  # set. If the whole mount vanished mid-page, every probe read as gone - skip
-  # the deletions rather than wipe a page of live rows.
-  @spec maybe_prune(t(), [Sweep.blob()]) ::
-          {non_neg_integer(), non_neg_integer(), non_neg_integer()}
-  defp maybe_prune(_state, []), do: {0, 0, 0}
-
-  defp maybe_prune(state, missing) do
-    case LayerRepo.test_system() do
-      :ok ->
-        prune_missing(state, missing)
-
-      {:error, reason} ->
-        Logger.warning(
-          "layer gc: medium became unavailable before pruning (#{inspect(reason)}); " <>
-            "skipping #{length(missing)} deletion(s) this page"
-        )
-
-        {0, 0, 0}
-    end
-  end
-
-  # Prune the missing blobs that no image references; report the rest as dangling.
-  @spec prune_missing(t(), [Sweep.blob()]) ::
-          {non_neg_integer(), non_neg_integer(), non_neg_integer()}
-  defp prune_missing(state, missing) do
-    ids = Enum.map(missing, fn {id, _size} -> id end)
-    referenced = referenced_ids(state, ids)
-
-    {dangling, prunable} =
-      Enum.split_with(missing, fn {id, _size} -> MapSet.member?(referenced, id) end)
-
-    report_dangling(dangling)
-
-    cutoff =
-      DateTime.add(DateTime.utc_now(), -Unit.Time.as_ms(state.config.grace_period), :millisecond)
-
-    prunable_ids = Enum.map(prunable, fn {id, _size} -> id end)
-    {pruned, pruned_bytes} = prune_rows(state, prunable_ids, cutoff)
-
-    {pruned, pruned_bytes, length(dangling)}
-  end
-
-  @spec prune_rows(t(), [String.t()], DateTime.t()) :: {non_neg_integer(), non_neg_integer()}
-  defp prune_rows(_state, [], _cutoff), do: {0, 0}
-
-  defp prune_rows(state, ids, cutoff) do
-    # `RETURNING size` gives the exact deleted set's count and bytes. The grace
-    # cutoff protects freshly-published rows. NOT EXISTS double-guards the FK:
-    # even if a reference appeared since we snapshotted `referenced`, that row is
-    # skipped rather than raising.
-    query =
-      from b in Blob,
-        as: :b,
-        where:
-          b.id in ^ids and b.state == :present and b.inserted_at < ^cutoff and
-            not exists(from il in ImageLayer, where: il.blob_id == parent_as(:b).id),
-        select: b.size
-
-    {count, sizes} = with_low_priority(state, fn -> Repo.delete_all(query) end)
-    {count, Enum.sum(sizes)}
-  end
-
-  # Report dangling blobs (file gone, still referenced = data loss) once per page,
-  # aggregated, so a broken mount cannot flood the logs one line per blob.
-  @spec report_dangling([Sweep.blob()]) :: :ok
-  defp report_dangling([]), do: :ok
-
-  defp report_dangling(dangling) do
-    count = length(dangling)
-    bytes = dangling |> Enum.map(fn {_id, size} -> size end) |> Enum.sum()
-    sample = dangling |> Enum.take(10) |> Enum.map(fn {id, _size} -> id end)
-
-    Logger.error(
-      "layer gc: #{count} blob(s) missing from medium but still referenced by an image " <>
-        "(#{bytes} bytes total); leaving in place. sample=#{inspect(sample)}"
-    )
-  end
-
-  @spec referenced_ids(t(), [String.t()]) :: MapSet.t(String.t())
-  defp referenced_ids(state, ids) do
-    query = from il in ImageLayer, where: il.blob_id in ^ids, distinct: true, select: il.blob_id
-    state |> with_low_priority(fn -> Repo.all(query) end) |> MapSet.new()
-  end
-
-  # Run a DB operation at low priority: in a transaction with a capped per-statement
-  # timeout, so it can never pin a backend and yields under contention.
-  @spec with_low_priority(t(), (-> result)) :: result when result: var
-  defp with_low_priority(state, fun) do
-    timeout = Unit.Time.as_ms(state.config.timeout)
-
-    {:ok, result} =
-      Repo.transaction(fn ->
-        _ =
-          Repo.query!("SELECT set_config('statement_timeout', $1, true)", [
-            Integer.to_string(timeout)
-          ])
-
-        fun.()
-      end)
-
-    result
-  end
-
-  # Shared-medium presence probe injected into the pure Sweep core. Distinguishes
-  # a genuine absence (`:enoent` -> prunable) from an I/O error (`:unknown` ->
-  # never pruned), so a transient NFS hiccup can never drive a delete.
-  @spec presence(String.t()) :: Sweep.presence()
-  defp presence(id) do
-    case LayerRepo.find_layer(id) do
-      {:ok, _path} -> :present
-      {:error, :enoent} -> :missing
-      {:error, _posix} -> :unknown
-    end
-  end
 end

lib/hyper/img/db/gc/prune.ex

@@ -0,0 +1,157 @@
+defmodule Hyper.Img.Db.Gc.Prune do
+  @moduledoc """
+  The delete step of the layer GC, factored out of the `Gc` server so its
+  safety contracts are testable against a real database without the
+  cluster-singleton GenServer around them:
+
+    * a missing blob still referenced by an image is *dangling* — reported,
+      never deleted;
+    * rows younger than `grace_period` are never deleted;
+    * the `DELETE` re-checks `NOT EXISTS (image_layers)`, so a reference
+      that appeared after the caller's snapshot is skipped, not violated;
+    * the shared medium is re-probed before any delete, so a mount that
+      vanished mid-page skips the page instead of mass-deleting;
+    * `presence/1` maps only a true `:enoent` to `:missing`; any other I/O
+      error is `:unknown` and never drives a delete.
+
+  Every query runs at low priority: in a transaction with a capped
+  per-statement timeout, so the GC can never pin a backend.
+  """
+
+  require Logger
+  import Ecto.Query
+
+  alias Hyper.Cfg.Gc, as: Config
+  alias Hyper.Img.Db.{Blob, ImageLayer, Repo}
+  alias Hyper.Img.Db.Gc.Sweep
+  alias Hyper.Node.Layer.Repo, as: LayerRepo
+
+  @doc """
+  Act on one page's missing set: re-check the medium is still mounted, split
+  dangling (still referenced) from prunable, delete what is out of grace.
+  Returns `{pruned, pruned_bytes, dangling}`.
+  """
+  @spec execute(Config.t(), [Sweep.blob()]) ::
+          {non_neg_integer(), non_neg_integer(), non_neg_integer()}
+  def execute(_config, []), do: {0, 0, 0}
+
+  def execute(config, missing) do
+    case LayerRepo.test_system() do
+      :ok ->
+        prune_missing(config, missing)
+
+      {:error, reason} ->
+        Logger.warning(
+          "layer gc: medium became unavailable before pruning (#{inspect(reason)}); " <>
+            "skipping #{length(missing)} deletion(s) this page"
+        )
+
+        {0, 0, 0}
+    end
+  end
+
+  @doc """
+  Shared-medium presence probe injected into the pure Sweep core.
+  Distinguishes a genuine absence (`:enoent` -> prunable) from an I/O error
+  (`:unknown` -> never pruned), so a transient NFS hiccup can never drive a
+  delete.
+  """
+  @spec presence(String.t()) :: Sweep.presence()
+  def presence(id) do
+    case LayerRepo.find_layer(id) do
+      {:ok, _path} -> :present
+      {:error, :enoent} -> :missing
+      {:error, _posix} -> :unknown
+    end
+  end
+
+  @doc """
+  Run a DB operation at low priority: in a transaction with a capped
+  per-statement timeout, so it can never pin a backend and yields under
+  contention.
+  """
+  @spec with_low_priority(Config.t(), (-> result)) :: result when result: var
+  def with_low_priority(config, fun) do
+    timeout = Unit.Time.as_ms(config.timeout)
+
+    {:ok, result} =
+      Repo.transaction(fn ->
+        _ =
+          Repo.query!("SELECT set_config('statement_timeout', $1, true)", [
+            Integer.to_string(timeout)
+          ])
+
+        fun.()
+      end)
+
+    result
+  end
+
+  @doc false
+  # Public only for its safety test: the NOT EXISTS re-check guards a race (a
+  # reference appearing after the caller snapshotted `referenced_ids`) that
+  # cannot be reached deterministically through `execute/2`.
+  @spec prune_rows(Config.t(), [String.t()], DateTime.t()) ::
+          {non_neg_integer(), non_neg_integer()}
+  def prune_rows(_config, [], _cutoff), do: {0, 0}
+
+  def prune_rows(config, ids, cutoff) do
+    # `RETURNING size` gives the exact deleted set's count and bytes. The grace
+    # cutoff protects freshly-published rows. NOT EXISTS double-guards the FK:
+    # even if a reference appeared since we snapshotted `referenced`, that row
+    # is skipped rather than raising.
+    query =
+      from b in Blob,
+        as: :b,
+        where:
+          b.id in ^ids and b.state == :present and b.inserted_at < ^cutoff and
+            not exists(from il in ImageLayer, where: il.blob_id == parent_as(:b).id),
+        select: b.size
+
+    {count, sizes} = with_low_priority(config, fn -> Repo.delete_all(query) end)
+    {count, Enum.sum(sizes)}
+  end
+
+  # Prune the missing blobs that no image references; report the rest as dangling.
+  @spec prune_missing(Config.t(), [Sweep.blob()]) ::
+          {non_neg_integer(), non_neg_integer(), non_neg_integer()}
+  defp prune_missing(config, missing) do
+    ids = Enum.map(missing, fn {id, _size} -> id end)
+    referenced = referenced_ids(config, ids)
+
+    {dangling, prunable} =
+      Enum.split_with(missing, fn {id, _size} -> MapSet.member?(referenced, id) end)
+
+    report_dangling(dangling)
+
+    cutoff =
+      DateTime.add(DateTime.utc_now(), -Unit.Time.as_ms(config.grace_period), :millisecond)
+
+    prunable_ids = Enum.map(prunable, fn {id, _size} -> id end)
+    {pruned, pruned_bytes} = prune_rows(config, prunable_ids, cutoff)
+
+    {pruned, pruned_bytes, length(dangling)}
+  end
+
+  # Report dangling blobs (file gone, still referenced = data loss) once per
+  # page, aggregated, so a broken mount cannot flood the logs one line per blob.
+  @spec report_dangling([Sweep.blob()]) :: :ok
+  defp report_dangling([]), do: :ok
+
+  defp report_dangling(dangling) do
+    count = length(dangling)
+    bytes = dangling |> Enum.map(fn {_id, size} -> size end) |> Enum.sum()
+    sample = dangling |> Enum.take(10) |> Enum.map(fn {id, _size} -> id end)
+
+    Logger.error(
+      "layer gc: #{count} blob(s) missing from medium but still referenced by an image " <>
+        "(#{bytes} bytes total); leaving in place. sample=#{inspect(sample)}"
+    )
+  end
+
+  @spec referenced_ids(Config.t(), [String.t()]) :: MapSet.t(String.t())
+  defp referenced_ids(config, ids) do
+    query = from il in ImageLayer, where: il.blob_id in ^ids, distinct: true, select: il.blob_id
+    config |> with_low_priority(fn -> Repo.all(query) end) |> MapSet.new()
+  end
+end