mirrors/garage
1
0
Fork 0
mirror of https://git.deuxfleurs.fr/Deuxfleurs/garage.git synced 2024-11-21 23:51:00 +00:00
Code Issues Projects Releases Packages Wiki Activity
garage/script/jepsen.garage/README.md
Alex Auvolat d2c365767b jepsen: more testing
2023-10-24 11:39:45 +02:00

6.6 KiB
Raw Blame History

jepsen.garage

Jepsen checking of Garage consistency properties.

Usage

Requirements:

  • vagrant
  • VirtualBox, configured so that nodes can take an IP in a private network 192.168.56.0/24
  • a user that can create VirtualBox VMs
  • leiningen
  • gnuplot

Set up VMs:

vagrant up

Run tests (this one should fail):

lein run test --nodes-file nodes.vagrant --time-limit 64 --concurrency 50 --rate 50 --workload reg

These ones are working:

lein run test --nodes-file nodes.vagrant --time-limit 64 --rate 50  --concurrency 50 --workload set1
lein run test --nodes-file nodes.vagrant --time-limit 64 --rate 50  --concurrency 50 --workload set2

Results

Register linear, without timestamp patch

Command: lein run test --nodes-file nodes.vagrant --time-limit 60 --rate 100 --concurrency 20 --workload reg1 --ops-per-key 100

Results without timestamp patch:

  • Fails with a simple clock-scramble nemesis (--scenario c). Explanation: without the timestamp patch, nodes will create objects using their local clock only as a timestamp, so the ordering will be all over the place if clocks are scrambled.

Results with timestamp patch (--patch tsfix2):

  • No failure with clock-scramble nemesis

  • Fails with clock-scramble nemesis + partition nemesis (--scenario cp).

This test is expected to fail. Indeed, S3 objects are not meant to behave like linearizable registers. TODO explain using a counter-example

Read-after-write CRDT register model

Command: lein run test --nodes-file nodes.vagrant --time-limit 60 --rate 100 --concurrency 100 --workload reg2 --ops-per-key 100

Results without timestamp patch:

  • Fails with a simple clock-scramble nemesis (--scenario c). Explanation: old values are not overwritten correctly when their timestamps are in the future.

Results with timestamp patch (--patch tsfix2):

  • No failures with clock-scramble nemesis + partition nemesis (--scenario cp). This proves that tsfix2 (PR#543) does improve consistency.

  • Fails with layout reconfiguration nemesis (--scenario r) (TODO: note down the run id of a failed run) (TODO: test more and investigate). This is the failure mode we are looking for and trying to fix for NLnet task 3.

Set, basic test (write some items, then read)

Command: lein run test --nodes-file nodes.vagrant --time-limit 60 --rate 100 --concurrency 100 --workload set1 --ops-per-key 100 --patch tsfix2

Results:

  • For now, no failures with clock-scramble nemesis + partition nemesis -> TODO long test run

  • Failures were not yet achieved with only the layout reconfiguration nemesis, although they should be.

  • Fails with partition + layout reconfiguration nemesis (--scenario pr) (TODO: note down the run id of a failed run) (TODO: test more and investigate). This is the failure mode we are looking for and trying to fix for NLnet task 3.

Set, continuous test (interspersed reads and writes)

Command: lein run test --nodes-file nodes.vagrant --time-limit 60 --rate 100 --concurrency 100 --workload set2 --ops-per-key 100 --patch tsfix2

Results:

  • For now, no failures with clock-scramble nemesis + partition nemesis -> TODO long test run

  • Failures were not yet achieved with only the layout reconfiguration nemesis, although they should be.

  • TODO: failures should be achieved with --scenario pr? Even with 4 or 5 consecutive test runs, no failures were achieved, why? (TODO: note down the run id of a failed run)

Investigating (and fixing) errors

Segfaults

They are due to the download being interrupted in the middle (^C during first launch on clean VMs), the garage binary is truncated. Add :force? to the cached-wget! call in daemon.clj to re-download the binary.

In jepsen.garage: prefix wierdness

In store/garage set1/20231019T163358.615+0200:

INFO [2023-10-19 16:35:20,977] clojure-agent-send-off-pool-207 - jepsen.garage.set list results for prefix set20/ : (set13/0 set13/1 set13/10 set13/11 set13/12 set13/13 set13/14 set13/15 set13/16 set13/17 set13/18 set13/19 set13/2 set13/20 set13/21 set13/22 set13/23 set13/24 set13/25 set13/26 set13/27 set13/28 set13/29 set13/3 set13/30 set13/31 set13/32 set13/33 set13/34 set13/35 set13/36 set13/37 set13/38 set13/39 set13/4 set13/40 set13/41 set13/42 set13/43 set13/44 set13/45 set13/46 set13/47 set13/48 set13/49 set13/5 set13/50 set13/51 set13/52 set13/53 set13/54 set13/55 set13/56 set13/57 set13/58 set13/59 set13/6 set13/60 set13/61 set13/62 set13/63 set13/64 set13/65 set13/66 set13/67 set13/68 set13/69 set13/7 set13/70 set13/71 set13/72 set13/73 set13/74 set13/75 set13/76 set13/77 set13/78 set13/79 set13/8 set13/80 set13/81 set13/82 set13/83 set13/84 set13/85 set13/86 set13/87 set13/88 set13/89 set13/9 set13/90 set13/91 set13/92 set13/93 set13/94 set13/95 set13/96 set13/97 set13/98 set13/99)  (node: http://192.168.56.25:3900 )

After inspecting, the actual S3 call made was with prefix "set13/", so at least this is not an error in Garage itself but in the jepsen code.

Finally found out that this was due to closures not correctly capturing their context in the list function in s3api.clj (wtf clojure?) Not sure exactly where it came from but it seems to have been fixed by making list-inner a separate function and not a sub-function, and passing all values that were previously in the context (creds and prefix) as additional arguments.

reg2 test inconsistency, even with timestamp fix

The reg2 test is our custom checker for CRDT read-after-write on individual object keys, acting as registers which can be updated. The test fails without the timestamp fix, which is expected as the clock scrambler will prevent nodes from having a correct ordering of objects.

With the timestamp fix (--patch tsfix1), the happenned-before relationship should at least be respected, meaning that when a PutObject call starts after another PutObject call has ended, the second call should overwrite the value of the first call, and that value should not be readable by future GetObject calls. However, we observed inconsistencies even with the timestamp fix.

The inconsistencies seemed to always happenned after writing a nil value, which translates to a DeleteObject call instead of a PutObject. By removing the possibility of writing nil values, therefore only doing PutObject calls, the issue disappears. There is therefore an issue to fix in DeleteObject.

The issue in DeleteObject seems to have been fixed by commit c82d91c6bccf307186332b6c5c6fc0b128b1b2b1, which can be used using --patch tsfix2.

License

Copyright © 2023 Alex Auvolat

This program and the accompanying materials are made available under the terms of the GNU Affero General Public License v3.0.