|
1
|
+ |
defmodule OpenAgents.EffectsTest do
|
|
2
|
+ |
@moduledoc """
|
|
3
|
+ |
What the durable effect outbox promises (EFFECT-001, EFFECT-002, issue #202).
|
|
4
|
+ |
|
|
5
|
+ |
The claim under test is not "effects usually run". It is that an effect
|
|
6
|
+ |
exists exactly when the intent that asked for it committed, that one worker
|
|
7
|
+ |
runs it at a time, that a worker that dies holding it loses nothing, and that
|
|
8
|
+ |
a redelivery is safe. Each of those is a separate failure the outbox exists
|
|
9
|
+ |
to remove, so each gets its own test.
|
|
10
|
+ |
"""
|
|
11
|
+ |
|
|
12
|
+ |
use OpenAgents.DataCase, async: false
|
|
13
|
+ |
|
|
14
|
+ |
alias OpenAgents.Effects
|
|
15
|
+ |
alias OpenAgents.Effects.Effect
|
|
16
|
+ |
alias OpenAgents.Effects.Worker
|
|
17
|
+ |
|
|
18
|
+ |
setup do
|
|
19
|
+ |
Application.put_env(:openagents, :effects,
|
|
20
|
+ |
handlers: %{"test.echo" => OpenAgents.EffectsEchoHandler},
|
|
21
|
+ |
backoff_base_ms: 1_000,
|
|
22
|
+ |
backoff_ceiling_ms: 300_000,
|
|
23
|
+ |
lease_seconds: 120
|
|
24
|
+ |
)
|
|
25
|
+ |
|
|
26
|
+ |
Application.put_env(:openagents, :effects_test_observer, self())
|
|
27
|
+ |
|
|
28
|
+ |
on_exit(fn ->
|
|
29
|
+ |
Application.delete_env(:openagents, :effects)
|
|
30
|
+ |
Application.delete_env(:openagents, :effects_test_observer)
|
|
31
|
+ |
end)
|
|
32
|
+ |
|
|
33
|
+ |
:ok
|
|
34
|
+ |
end
|
|
35
|
+ |
|
|
36
|
+ |
describe "enqueue/2 inside the caller's transaction" do
|
|
37
|
+ |
test "a committed transaction leaves exactly one effect" do
|
|
38
|
+ |
{:ok, effect} =
|
|
39
|
+ |
Repo.transaction(fn ->
|
|
40
|
+ |
{:ok, effect} = enqueue("commit-me")
|
|
41
|
+ |
effect
|
|
42
|
+ |
end)
|
|
43
|
+ |
|
|
44
|
+ |
assert %Effect{status: "pending", attempts: 0} = Effects.get(effect.id)
|
|
45
|
+ |
assert Effects.counts() == %{"pending" => 1}
|
|
46
|
+ |
end
|
|
47
|
+ |
|
|
48
|
+ |
test "a rolled-back transaction leaves no effect at all" do
|
|
49
|
+ |
key = Effects.idempotency_key("test.echo", "test_source", "rollback-me")
|
|
50
|
+ |
|
|
51
|
+ |
assert {:error, :intent_refused} =
|
|
52
|
+ |
Repo.transaction(fn ->
|
|
53
|
+ |
{:ok, _effect} = enqueue("rollback-me")
|
|
54
|
+ |
Repo.rollback(:intent_refused)
|
|
55
|
+ |
end)
|
|
56
|
+ |
|
|
57
|
+ |
# This is the whole point of enqueuing inside the caller's transaction:
|
|
58
|
+ |
# an intent that did not happen owes nothing, and nothing is delivered.
|
|
59
|
+ |
assert Effects.get_by_key(key) == nil
|
|
60
|
+ |
assert Effects.counts() == %{}
|
|
61
|
+ |
end
|
|
62
|
+ |
|
|
63
|
+ |
test "the same intent enqueued twice is one effect and one delivery" do
|
|
64
|
+ |
{:ok, first} = enqueue("twice")
|
|
65
|
+ |
{:ok, second} = enqueue("twice")
|
|
66
|
+ |
|
|
67
|
+ |
assert first.id == second.id
|
|
68
|
+ |
assert Repo.aggregate(Effect, :count) == 1
|
|
69
|
+ |
end
|
|
70
|
+ |
|
|
71
|
+ |
test "a reused key carrying different content is refused, not silently answered" do
|
|
72
|
+ |
{:ok, first} = enqueue("fingerprinted", %{"body" => "original"})
|
|
73
|
+ |
|
|
74
|
+ |
assert {:error, :payload_conflict} =
|
|
75
|
+ |
Effects.enqueue("test.echo", %{
|
|
76
|
+ |
payload: %{"body" => "substituted"},
|
|
77
|
+ |
source_kind: "test_source",
|
|
78
|
+ |
source_id: "fingerprinted"
|
|
79
|
+ |
})
|
|
80
|
+ |
|
|
81
|
+ |
# The first caller's effect stands; the second caller is told no rather
|
|
82
|
+ |
# than handed a result for a payload it never sent.
|
|
83
|
+ |
assert Effects.get(first.id).payload == %{"body" => "original"}
|
|
84
|
+ |
assert Repo.aggregate(Effect, :count) == 1
|
|
85
|
+ |
end
|
|
86
|
+ |
|
|
87
|
+ |
test "the deterministic key does not depend on the payload" do
|
|
88
|
+ |
key = Effects.idempotency_key("test.echo", "test_source", "stable", 7)
|
|
89
|
+ |
|
|
90
|
+ |
assert key == Effects.idempotency_key("test.echo", "test_source", "stable", 7)
|
|
91
|
+ |
refute key == Effects.idempotency_key("test.echo", "test_source", "stable")
|
|
92
|
+ |
refute key == Effects.idempotency_key("test.other", "test_source", "stable", 7)
|
|
93
|
+ |
end
|
|
94
|
+ |
|
|
95
|
+ |
test "a source sequence is recorded as evidence, never as a status" do
|
|
96
|
+ |
{:ok, effect} =
|
|
97
|
+ |
Effects.enqueue("test.echo", %{
|
|
98
|
+ |
payload: %{"body" => "sequenced"},
|
|
99
|
+ |
source_kind: "thread_event",
|
|
100
|
+ |
source_id: "thread-1",
|
|
101
|
+ |
source_sequence: 42
|
|
102
|
+ |
})
|
|
103
|
+ |
|
|
104
|
+ |
# EFFECT-002: a transcript position is not an execution claim and not a
|
|
105
|
+ |
# completion claim. The sequence is stored; the status is separate.
|
|
106
|
+ |
assert effect.source_sequence == 42
|
|
107
|
+ |
assert effect.status == "pending"
|
|
108
|
+ |
assert effect.claimed_at == nil
|
|
109
|
+ |
assert effect.completed_at == nil
|
|
110
|
+ |
end
|
|
111
|
+ |
end
|
|
112
|
+ |
|
|
113
|
+ |
describe "claim_batch/2" do
|
|
114
|
+ |
test "a claim takes a lease and counts an attempt" do
|
|
115
|
+ |
{:ok, effect} = enqueue("claim-me")
|
|
116
|
+ |
|
|
117
|
+ |
assert [claimed] = Effects.claim_batch("worker-a")
|
|
118
|
+ |
assert claimed.id == effect.id
|
|
119
|
+ |
assert claimed.status == "claimed"
|
|
120
|
+ |
assert claimed.attempts == 1
|
|
121
|
+ |
assert claimed.lease_owner == "worker-a"
|
|
122
|
+ |
assert DateTime.compare(claimed.lease_expires_at, DateTime.utc_now()) == :gt
|
|
123
|
+ |
|
|
124
|
+ |
# Claiming is not completing (EFFECT-002).
|
|
125
|
+ |
assert claimed.claimed_at != nil
|
|
126
|
+ |
assert claimed.completed_at == nil
|
|
127
|
+ |
end
|
|
128
|
+ |
|
|
129
|
+ |
test "an effect a worker holds is not offered to the next worker" do
|
|
130
|
+ |
{:ok, _effect} = enqueue("held")
|
|
131
|
+ |
|
|
132
|
+ |
assert [_claimed] = Effects.claim_batch("worker-a")
|
|
133
|
+ |
assert Effects.claim_batch("worker-b") == []
|
|
134
|
+ |
end
|
|
135
|
+ |
|
|
136
|
+ |
test "an effect whose time has not come is not claimable" do
|
|
137
|
+ |
later = DateTime.add(DateTime.utc_now(), 60, :second)
|
|
138
|
+ |
|
|
139
|
+ |
{:ok, _effect} =
|
|
140
|
+ |
Effects.enqueue("test.echo", enqueue_attributes("later", available_at: later))
|
|
141
|
+ |
|
|
142
|
+ |
assert Effects.claim_batch("worker-a") == []
|
|
143
|
+ |
assert [_claimed] = Effects.claim_batch("worker-a", now: DateTime.add(later, 1, :second))
|
|
144
|
+ |
end
|
|
145
|
+ |
|
|
146
|
+ |
test "concurrent workers over one batch claim disjoint sets and never the same effect twice" do
|
|
147
|
+ |
for index <- 1..12, do: {:ok, _effect} = enqueue("racer-#{index}")
|
|
148
|
+ |
|
|
149
|
+ |
claims =
|
|
150
|
+ |
["worker-a", "worker-b", "worker-c"]
|
|
151
|
+ |
|> Task.async_stream(
|
|
152
|
+ |
fn worker -> Effects.claim_batch(worker, limit: 12) end,
|
|
153
|
+ |
max_concurrency: 3,
|
|
154
|
+ |
ordered: false,
|
|
155
|
+ |
timeout: :infinity
|
|
156
|
+ |
)
|
|
157
|
+ |
|> Enum.flat_map(fn {:ok, claimed} -> claimed end)
|
|
158
|
+ |
|
|
159
|
+ |
ids = Enum.map(claims, & &1.id)
|
|
160
|
+ |
|
|
161
|
+ |
# Every effect went to exactly one worker: no effect is missing, and no
|
|
162
|
+ |
# effect was handed to two workers to run twice.
|
|
163
|
+ |
assert length(ids) == 12
|
|
164
|
+ |
assert length(Enum.uniq(ids)) == 12
|
|
165
|
+ |
assert Enum.all?(claims, &(&1.attempts == 1))
|
|
166
|
+ |
assert Effects.counts() == %{"claimed" => 12}
|
|
167
|
+ |
end
|
|
168
|
+ |
|
|
169
|
+ |
test "a claim only offers kinds this release can run" do
|
|
170
|
+ |
{:ok, _known} = enqueue("known")
|
|
171
|
+ |
|
|
172
|
+ |
{:ok, _unknown} =
|
|
173
|
+ |
Effects.enqueue("test.absent", enqueue_attributes("unknown"))
|
|
174
|
+ |
|
|
175
|
+ |
assert [claimed] = Effects.claim_batch("worker-a", kinds: ["test.echo"])
|
|
176
|
+ |
assert claimed.kind == "test.echo"
|
|
177
|
+ |
end
|
|
178
|
+ |
end
|
|
179
|
+ |
|
|
180
|
+ |
describe "reclaim_expired/1" do
|
|
181
|
+ |
test "a dead worker's lease returns the effect to the queue" do
|
|
182
|
+ |
{:ok, _effect} = enqueue("abandoned")
|
|
183
|
+ |
assert [claimed] = Effects.claim_batch("worker-a", lease_seconds: 1)
|
|
184
|
+ |
|
|
185
|
+ |
after_expiry = DateTime.add(claimed.lease_expires_at, 1, :second)
|
|
186
|
+ |
|
|
187
|
+ |
assert Effects.reclaim_expired(now: after_expiry) == 1
|
|
188
|
+ |
|
|
189
|
+ |
reclaimed = Effects.get(claimed.id)
|
|
190
|
+ |
assert reclaimed.status == "pending"
|
|
191
|
+ |
assert reclaimed.lease_owner == nil
|
|
192
|
+ |
assert reclaimed.lease_expires_at == nil
|
|
193
|
+ |
|
|
194
|
+ |
# The attempt the dead worker spent is not refunded, so a handler that
|
|
195
|
+ |
# reliably kills its worker still reaches maximum_attempts and stops.
|
|
196
|
+ |
assert reclaimed.attempts == 1
|
|
197
|
+ |
|
|
198
|
+ |
assert [reclaimed_again] = Effects.claim_batch("worker-b", now: after_expiry)
|
|
199
|
+ |
assert reclaimed_again.lease_owner == "worker-b"
|
|
200
|
+ |
assert reclaimed_again.attempts == 2
|
|
201
|
+ |
end
|
|
202
|
+ |
|
|
203
|
+ |
test "a live lease is left alone" do
|
|
204
|
+ |
{:ok, _effect} = enqueue("live")
|
|
205
|
+ |
assert [claimed] = Effects.claim_batch("worker-a", lease_seconds: 600)
|
|
206
|
+ |
|
|
207
|
+ |
assert Effects.reclaim_expired() == 0
|
|
208
|
+ |
assert Effects.get(claimed.id).lease_owner == "worker-a"
|
|
209
|
+ |
end
|
|
210
|
+ |
end
|
|
211
|
+ |
|
|
212
|
+ |
describe "fail/2" do
|
|
213
|
+ |
test "a failure backs off, releases the lease, and is retried" do
|
|
214
|
+ |
{:ok, _effect} = enqueue("flaky")
|
|
215
|
+ |
assert [claimed] = Effects.claim_batch("worker-a")
|
|
216
|
+ |
|
|
217
|
+ |
before = DateTime.utc_now()
|
|
218
|
+ |
assert {:ok, failed} = Effects.fail(claimed, {:provider_unavailable, 503})
|
|
219
|
+ |
|
|
220
|
+ |
assert failed.status == "pending"
|
|
221
|
+ |
assert failed.lease_owner == nil
|
|
222
|
+ |
assert failed.last_error =~ "provider_unavailable"
|
|
223
|
+ |
assert failed.attempts == 1
|
|
224
|
+ |
|
|
225
|
+ |
# Backoff is a delay, not a refusal: the effect is deliverable again once
|
|
226
|
+ |
# its time comes, and not before.
|
|
227
|
+ |
assert DateTime.diff(failed.available_at, before, :millisecond) >= Effects.backoff_ms(1)
|
|
228
|
+ |
assert Effects.claim_batch("worker-b") == []
|
|
229
|
+ |
|
|
230
|
+ |
later = DateTime.add(failed.available_at, 1, :second)
|
|
231
|
+ |
assert [retried] = Effects.claim_batch("worker-b", now: later)
|
|
232
|
+ |
assert retried.attempts == 2
|
|
233
|
+ |
end
|
|
234
|
+ |
|
|
235
|
+ |
test "backoff grows and is capped" do
|
|
236
|
+ |
assert Effects.backoff_ms(1) == 1_000
|
|
237
|
+ |
assert Effects.backoff_ms(2) == 2_000
|
|
238
|
+ |
assert Effects.backoff_ms(3) == 4_000
|
|
239
|
+ |
assert Effects.backoff_ms(40) == 300_000
|
|
240
|
+ |
end
|
|
241
|
+ |
|
|
242
|
+ |
test "an effect that exhausts its attempts stops being delivered" do
|
|
243
|
+ |
{:ok, _effect} =
|
|
244
|
+ |
Effects.enqueue("test.echo", enqueue_attributes("doomed", maximum_attempts: 2))
|
|
245
|
+ |
|
|
246
|
+ |
assert [first] = Effects.claim_batch("worker-a")
|
|
247
|
+ |
assert {:ok, retryable} = Effects.fail(first, :first_failure)
|
|
248
|
+ |
assert retryable.status == "pending"
|
|
249
|
+ |
|
|
250
|
+ |
later = DateTime.add(retryable.available_at, 1, :second)
|
|
251
|
+ |
assert [second] = Effects.claim_batch("worker-a", now: later)
|
|
252
|
+ |
assert second.attempts == 2
|
|
253
|
+ |
|
|
254
|
+ |
assert {:ok, dead} = Effects.fail(second, :second_failure)
|
|
255
|
+ |
|
|
256
|
+ |
# An effect that cannot be run must stop pretending it will be, so that
|
|
257
|
+ |
# something else can notice it.
|
|
258
|
+ |
assert dead.status == "failed"
|
|
259
|
+ |
assert dead.completed_at != nil
|
|
260
|
+ |
assert dead.lease_owner == nil
|
|
261
|
+ |
assert Effects.claim_batch("worker-a", now: DateTime.add(later, 3_600, :second)) == []
|
|
262
|
+ |
end
|
|
263
|
+ |
end
|
|
264
|
+ |
|
|
265
|
+ |
describe "complete/1" do
|
|
266
|
+ |
test "completion is idempotent under redelivery" do
|
|
267
|
+ |
{:ok, _effect} = enqueue("redelivered")
|
|
268
|
+ |
assert [claimed] = Effects.claim_batch("worker-a")
|
|
269
|
+ |
|
|
270
|
+ |
assert {:ok, done} = Effects.complete(claimed)
|
|
271
|
+ |
assert done.status == "done"
|
|
272
|
+ |
assert done.completed_at != nil
|
|
273
|
+ |
assert done.lease_owner == nil
|
|
274
|
+ |
|
|
275
|
+ |
# The second worker — the one whose lease expired mid-flight and whose
|
|
276
|
+ |
# effect someone else already finished — reports success without
|
|
277
|
+ |
# contradicting the record or writing a second completion.
|
|
278
|
+ |
assert {:ok, again} = Effects.complete(claimed)
|
|
279
|
+ |
assert again.id == done.id
|
|
280
|
+ |
assert again.status == "done"
|
|
281
|
+ |
assert again.completed_at == done.completed_at
|
|
282
|
+ |
end
|
|
283
|
+ |
|
|
284
|
+ |
test "completing a failed effect after the fact does not resurrect a failure" do
|
|
285
|
+ |
{:ok, _effect} = enqueue("late")
|
|
286
|
+ |
assert [claimed] = Effects.claim_batch("worker-a")
|
|
287
|
+ |
assert {:ok, _done} = Effects.complete(claimed)
|
|
288
|
+ |
|
|
289
|
+ |
# A stale worker reporting failure for an effect already completed does
|
|
290
|
+ |
# not turn a completed effect back into pending work.
|
|
291
|
+ |
assert {:ok, unchanged} = Effects.fail(claimed, :too_late)
|
|
292
|
+ |
assert unchanged.status == "done"
|
|
293
|
+ |
end
|
|
294
|
+ |
end
|
|
295
|
+ |
|
|
296
|
+ |
describe "the worker" do
|
|
297
|
+ |
test "one pass claims, dispatches, and completes" do
|
|
298
|
+ |
{:ok, effect} = enqueue("dispatch-me", %{"body" => "hello"})
|
|
299
|
+ |
|
|
300
|
+ |
assert %{claimed: 1, completed: 1, failed: 0} = Worker.run_once(identity: "worker-a")
|
|
301
|
+ |
|
|
302
|
+ |
assert_received {:effect_ran, "hello", key, id}
|
|
303
|
+ |
assert key == effect.idempotency_key
|
|
304
|
+ |
assert id == effect.id
|
|
305
|
+ |
assert Effects.get(effect.id).status == "done"
|
|
306
|
+ |
end
|
|
307
|
+ |
|
|
308
|
+ |
test "a handler that raises is a retry, not a crash" do
|
|
309
|
+ |
{:ok, effect} = enqueue("boom", %{"raise" => "handler exploded"})
|
|
310
|
+ |
|
|
311
|
+ |
assert %{claimed: 1, completed: 0, failed: 1} = Worker.run_once(identity: "worker-a")
|
|
312
|
+ |
|
|
313
|
+ |
failed = Effects.get(effect.id)
|
|
314
|
+ |
assert failed.status == "pending"
|
|
315
|
+ |
assert failed.last_error =~ "handler exploded"
|
|
316
|
+ |
assert failed.attempts == 1
|
|
317
|
+ |
end
|
|
318
|
+ |
|
|
319
|
+ |
test "a pass reclaims expired leases before it claims" do
|
|
320
|
+ |
{:ok, effect} = enqueue("stranded", %{"body" => "recovered"})
|
|
321
|
+ |
assert [claimed] = Effects.claim_batch("dead-worker", lease_seconds: -1)
|
|
322
|
+ |
assert claimed.status == "claimed"
|
|
323
|
+ |
|
|
324
|
+ |
assert %{reclaimed: 1, claimed: 1, completed: 1} = Worker.run_once(identity: "worker-b")
|
|
325
|
+ |
|
|
326
|
+ |
assert_received {:effect_ran, "recovered", _key, _id}
|
|
327
|
+ |
assert Effects.get(effect.id).status == "done"
|
|
328
|
+ |
end
|
|
329
|
+ |
|
|
330
|
+ |
test "an effect whose kind has no handler fails loudly rather than vanishing" do
|
|
331
|
+ |
{:ok, effect} = Effects.enqueue("test.absent", enqueue_attributes("orphan"))
|
|
332
|
+ |
|
|
333
|
+ |
# The claim only offers admitted kinds, so an unregistered kind is never
|
|
334
|
+ |
# picked up and quietly marked done.
|
|
335
|
+ |
assert %{claimed: 0} = Worker.run_once(identity: "worker-a")
|
|
336
|
+ |
assert Effects.get(effect.id).status == "pending"
|
|
337
|
+ |
|
|
338
|
+ |
# Dispatched directly — as a recovery path would — it is a refusal.
|
|
339
|
+ |
assert [claimed] = Effects.claim_batch("worker-a")
|
|
340
|
+ |
assert :error = Worker.dispatch(claimed)
|
|
341
|
+ |
assert Effects.get(effect.id).last_error =~ "unknown_kind"
|
|
342
|
+ |
end
|
|
343
|
+ |
|
|
344
|
+ |
test "a running worker drives a pass on demand, with no sleeping" do
|
|
345
|
+ |
{:ok, effect} = enqueue("ticked", %{"body" => "tick"})
|
|
346
|
+ |
|
|
347
|
+ |
worker =
|
|
348
|
+ |
start_supervised!({Worker, name: :effects_test_worker, poll: false, identity: "worker-t"})
|
|
349
|
+ |
|
|
350
|
+ |
assert %{claimed: 1, completed: 1} = Worker.tick(worker)
|
|
351
|
+ |
assert_received {:effect_ran, "tick", _key, _id}
|
|
352
|
+ |
assert Effects.get(effect.id).status == "done"
|
|
353
|
+ |
end
|
|
354
|
+ |
end
|
|
355
|
+ |
|
|
356
|
+ |
describe "for_source/2" do
|
|
357
|
+ |
test "an intent can be asked what it is owed" do
|
|
358
|
+ |
{:ok, first} = enqueue("audited-1")
|
|
359
|
+ |
{:ok, second} = enqueue("audited-2")
|
|
360
|
+ |
|
|
361
|
+ |
assert Effects.for_source("test_source", "audited-1") |> Enum.map(& &1.id) == [first.id]
|
|
362
|
+ |
assert Effects.for_source("test_source", "audited-2") |> Enum.map(& &1.id) == [second.id]
|
|
363
|
+ |
assert Effects.for_source("test_source", "never-asked") == []
|
|
364
|
+ |
end
|
|
365
|
+ |
end
|
|
366
|
+ |
|
|
367
|
+ |
defp enqueue(source_id, payload \\ %{"body" => "noop"}) do
|
|
368
|
+ |
Effects.enqueue("test.echo", enqueue_attributes(source_id, payload: payload))
|
|
369
|
+ |
end
|
|
370
|
+ |
|
|
371
|
+ |
defp enqueue_attributes(source_id, extra \\ []) do
|
|
372
|
+ |
[payload: %{"body" => "noop"}, source_kind: "test_source", source_id: source_id]
|
|
373
|
+ |
|> Keyword.merge(extra)
|
|
374
|
+ |
|> Map.new()
|
|
375
|
+ |
end
|
|
376
|
+ |
end
|