Test your filesystem's ability to correctly serialize writes as expected in APPEND mode on a POSIX system.
mikes-air:test-atomic-writes mike$ ./test-atomic-writes -s 1
Between 1 and 222 workers, writing at least 1 each, of at least size 2 required to test.
Usage of ./test-atomic-writes:
-c int
c(ount): of writes per worker, use at least 1 (default 50)
-f string
f(ilename) to use for test (default "out.tmp")
-readonly
readonly: just run the validate function, use identical flags
-s int
s(size): in bytes to write, use at least 2 (default 4096)
-w int
w(orkers): number of concurent writers, use at least 1 and no more than 222 (default 50)
-worker int
worker: perform the writes, internal (default -1)
Note: The readonly switch also needs the identical parameters used to create the file.
go test -v main_test.go
cd api
go test -v start.go start_test.go
- See the test results from the last deployment at https://test-atomic-writes.vercel.app/test-output.log.
- See the build output from the last deployment at https://test-atomic-writes.vercel.app/build-output.log.
- Fix api/start to call the test code properly. Currently just demonstrated by extracting go environment info and executing "ls" and "pwd".
- Write in append mode in Mac OS X APFS doesn't appear to be atomic. Resolved 6.9.2018.
- Redeploying to vercel w/ Github integration.
- Adjusting build.sh.
- Lamda code is running and producing output at https://test-atomic-writes.vercel.app/api/start.
- Able to deploy to Vercel to build and test. Demonstrates a lamba API. Added required build.sh.
- Test both parts by using:
go test -v . ./api
- Continuing to work on getting this go function running on Vercel. Code compiles and runs locally. Fails to build:
2020-08-15T19:49:14.438Z Error: Could not find an exported function in "api/test-atomic-writes.go"
2020-08-15T19:49:14.438Z Learn more: https://vercel.com/docs/runtimes#official-runtimes/go
- Added a test for the API to force it to compile with out having to deploy.
- Days of wfh and physical isolation: 148
- Zeit deco'd version 1.0. So this is first forray into Vercel. Will attempt to deploy this go cmdline program as a serverless function.
- The idea is to make an API that will exec that cmdline app and direct it's output as the http respons.
- Replaced now with vercel. Go API is working to replace DOCKER setup. This is better because the test can be triggered by querring the API.
- Added deployment details for Zeit Now v1 where the default test ran successfully.
- Tested new beta for 10.13.5:
mikes-air:test-atomic-write mike$ uname -a
Darwin mikes-air.local 17.7.0 Darwin Kernel Version 17.7.0: Tue Jun 12 21:39:16 PDT 2018; root:xnu-4570.70.24~11/RELEASE_X86_64 x86_64
mikes-air:test-atomic-write mike$ ./test-atomic-write -s 1500 -w 222 -c 1208
Each line of file out.tmp will be 1500 bytes, written by 222 workers, writing 1208 lines each.
validate: source changes=261766 shuffle=1179.13 msgCnt=268176 errCnt=0
mikes-air:test-atomic-write mike$ ./test-atomic-write -s 15003 -w 222 -c 1208
Each line of file out.tmp will be 15003 bytes, written by 222 workers, writing 1208 lines each.
validate: source changes=264756 shuffle=1192.59 msgCnt=268176 errCnt=0
- Apple asked me to test 10.14, it's not out for public beta so I retested to get ready. Apfs seems to be working on High Sierra 10.13.5.
mikes-air:test-atomic-write mike$ uname -a
Darwin mikes-air.local 17.6.0 Darwin Kernel Version 17.6.0: Tue May 8 15:22:16 PDT 2018; root:xnu-4570.61.1~1/RELEASE_X86_64 x86_64
mikes-air:test-atomic-write mike$ ./test-atomic-write -s 15913 -w 100 -n 1200
Each line of file out.tmp will be 15913 bytes, written by 100 workers, writing 1200 lines each.
validate: source changes=119235 shuffle=1192.35 msgCnt=120000 errCnt=0
mikes-air:test-atomic-write mike$ df -T apfs
Filesystem 512-blocks Used Available Capacity iused ifree %iused Mounted on
/dev/disk1s1 489825072 242941256 216193168 53% 1504897 9223372036853270910 0% /
/dev/disk1s4 489825072 4194344 216193168 2% 3 9223372036854775804 0% /private/var/vm
/dev/disk1s5 489825072 24114992 216193168 11% 432634 9223372036854343173 0% /Volumes/mohave
mikes-air:test-atomic-write mike$
- Published to Github
- Reported Apple Bug #37859698
- MacOS APFS Volume • APFS (Encrypted) -- Fails
mikes-air:test-atomic-writes mike$ ./test-atomic-writes -n 100
Each line of file out.tmp will be 4096 bytes, written by 50 workers, writing 100 lines each.
validate: filesize missmatch: filesize=20467712 expected file size=20480000
validate: missed messages: msgCnt=4997 expected=5000 difference=3
validate: source changes=60 shuffle=1.20 msgCnt=4997 errCnt=2
mikes-air:test-atomic-writes mike$ df -T apfs
Filesystem 512-blocks Used Available Capacity iused ifree %iused Mounted on
/dev/disk1s1 489825072 220048920 262175048 46% 1468208 9223372036853307599 0% /
/dev/disk1s4 489825072 6291496 262175048 3% 3 9223372036854775804 0% /private/var/vm
- MAcOS USB External Physical Volume • Mac OS Extended (Journaled) -- Works
- AWS FreeBSD Local Volume • ufs -- Works
- AWS Amazon-Linux Local Volume • ext4 -- Works
Amazon_Linux-512MB-Ohio-1
18.219.2.238
[ec2-user@ip-172-26-2-243 test-atomic-write]$ ./test-atomic-write -n 1000 -s 15013 -w 100
Each line of out.tmp will be 15013 characters long, writen by 100 workers writing 1000 lines each.
validate: fd=3 filesize=1501300000 efs=1501300000
validate: source changes=636 shuffle=6.36 msgCnt=100000 errCnt=0
[ec2-user@ip-172-26-2-243]$ df -T
Filesystem Type 1K-blocks Used Available Use% Mounted on
devtmpfs devtmpfs 241544 56 241488 1% /dev
tmpfs tmpfs 250508 0 250508 0% /dev/shm
/dev/xvda1 ext4 20509288 3111320 17297720 16% /
[ec2-user@ip-172-26-2-243 ~]$
FreeBSD-512MB-Ohio-1
18.219.169.234
[ec2-user@ip-172-26-8-133 /usr/home/ec2-user/src/test-atomic-write]$ ./test-atomic-write -n 1000 -s 15913 -w 100
Each line of file out.tmp will be 15913 bytes, writen by 100 workers, writing 1000 lines each.
validate: source changes=86299 shuffle=862.99 msgCnt=100000 errCnt=0
[ec2-user@ip-172-26-8-133 /usr/home/ec2-user/src/test-atomic-write]$
- Tested on AWS lightsail linux. Append was unfailing at all sizes and counts up to 16k.
- Tested on AWS FreeBSD lightsail too.
mikes-air:test-atomic-write mike$ uname -a
Darwin mikes-air.local 17.4.0 Darwin Kernel Version 17.4.0: Sun Dec 17 09:19:54 PST 2017; root:xnu-4570.41.2~1/RELEASE_X86_64 x86_64
dtruss output.
open("out.tmp\0", 0x20A, 0x180) = 3 0
- Created FreeBSD linux
[ec2-user@ip-172-26-8-133 /usr/home/ec2-user/src/test-atomic-write]$ uname -a
FreeBSD ip-172-26-8-133 11.1-RELEASE-p4 FreeBSD 11.1-RELEASE-p4 #0: Tue Nov 14 06:12:40 UTC 2017 [email protected]:/usr/obj/usr/src/sys/GENERIC amd64
[ec2-user@ip-172-26-8-133 /usr/home/ec2-user/src/test-atomic-write]$
[ec2-user@ip-172-26-8-133 /usr/home/ec2-user/src/test-atomic-write]$ ./test-atomic-write -s 500013 -w 100
Each line of file out.tmp will be 500013 bytes, writen by 100 workers, writing 50 lines each.
validate: source changes=4745 shuffle=47.45 msgCnt=5000 errCnt=0
[ec2-user@ip-172-26-8-133 /usr/home/ec2-user/src/test-atomic-write]$
- Created a lightsail instance on AWS to compare test-atomic-write results between MacOS and Linux.
Amazon_Linux-512MB-Ohio-1
18.219.2.238
[ec2-user@ip-172-26-2-243 test-atomic-write]$ uname -a
Linux ip-172-26-2-243 4.9.76-3.78.amzn1.x86_64 #1 SMP Fri Jan 12 19:51:35 UTC 2018 x86_64 x86_64 x86_64 GNU/Linux
[ec2-user@ip-172-26-2-243 test-atomic-write]$ ./test-atomic-write -n 10000 -s 16000 &
[1]+ Running ./test-atomic-write -n 10000 -s 16000 &
Each line of out.tmp will be 16000 characters long, writen by 50 workers writing 10000 lines each.
[ec2-user@ip-172-26-2-243 test-atomic-write]$ validate: fd=3 filesize=800000000 efs=800000000
[ec2-user@ip-172-26-2-243 test-atomic-write]$ validate: source changes=3431 shuffle=68.62 msgCnt=500000 errCnt=0
[1]+ Done ./test-atomic-write -n 10000 -s 16000
- Cleaned up stats. test works, but results seem to suggest that whenever the writes are shuffled between writers, writes are lost.
- writing an append write test program. Modeled after Oz Soloman's shell script for the same purpose. His can be found here.