Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

rocksdb log using up the diskspace on flash #519

Merged
merged 1 commit into from
Dec 7, 2022

Conversation

paulmchen
Copy link
Contributor

This issue is caused by the default max_total_wal_size rocksdb setting used. The default takes 0 as the value for this setting, RocksDB will dynamically choose the WAL size limit to be calculated using the following formula:

[sum of all write_buffer_size * max_write_buffer_number] * 4].

In the specific test case done by us with a long running flash test, it does a large number of Writes/Deletes operations for 5 days, the number of write buffer is approx 2000, therefore the value (using the default max_total_wal_size) = 64 * 2000 * 4 = 500 GB. It causes the force flush may not happen until 500 GB diskspace reaches. Since we have a 128 GB SSD disk, the server will fail due to "running out of the disk space on the SSD".

The problem can be identified by going to our flush data path (e.g. /root/data), and run "ls -l *.log |wc -l", we can see a large number of *.log files there and each log is approx. 50 MB in size. If we add all these log files, the total diskspaces used by these logs are about 120GB. That is why it crashed the server with "running out of the diskspaces".

This fix is to use the settings for max_total_wal_size as 1 GB recommended by the RocksDB community, then i have no longer seeing the same issue with running exactly the same long running test cases.

Here is the test case that we ran to verified with the fix:

#!/bin/bash

for ((i=0;i<200;i++));
do
echo "#$i set"
./memtier_benchmark -s 172.31.14.101 -p 6379 -t 10 -c 100 -n 50000 --command='SET key data' --data-size-range=1-32 --key-minimum=1 --key-maximum=50000000 --command-key-pattern=P
echo "$i get"
./memtier_benchmark -s 172.31.14.101 -p 6379 -t 10 -c 100 -n 50000 --command='GET key' --key-minimum=1 --key-maximum=50000000 --command-key-pattern=P
echo "$i del"
./memtier_benchmark -s 172.31.14.101 -p 6379 -t 10 -c 100 -n 50000 --command='DEL key' --key-minimum=1 --key-maximum=50000000 --command-key-pattern=P
sleep 120
done

@msotheeswaran-sc msotheeswaran-sc merged commit fad2e87 into Snapchat:main Dec 7, 2022
msotheeswaran-sc added a commit that referenced this pull request Dec 14, 2022
* need to include stdint for uintptr_t

* need to include stdint for uintptr_t

* use atomic_load for g_pserver->mstime

* use atomic_load for g_pserver->mstime

* Integrate readwritelock with Pro Code

* Integrate readwritelock with Pro Code

* Defensive asserts for RWLock

* Defensive asserts for RWLock

* Save and restore master info in rdb to allow active replica partial sync (#371)

* save replid for all masters in rdb

* expanded rdbSaveInfo to hold multiple master structs

* parse repl-masters from rdb

* recover replid info from rdb in active replica mode, attempt partial sync

* save offset from rdb into correct variable

* don't change replid based on master in active rep

* save and load psync info from correct fields

* Save and restore master info in rdb to allow active replica partial sync (#371)

* save replid for all masters in rdb

* expanded rdbSaveInfo to hold multiple master structs

* parse repl-masters from rdb

* recover replid info from rdb in active replica mode, attempt partial sync

* save offset from rdb into correct variable

* don't change replid based on master in active rep

* save and load psync info from correct fields

* placement new instead of memcpy

* placement new instead of memcpy

* Remove asserts, RW lock can go below zero in cases of aeAcquireLock

* Remove asserts, RW lock can go below zero in cases of aeAcquireLock

* Inclusive language

* Inclusive language

* update packaging for OS merge

* update packaging for OS merge

* modify dockerfile to build within image

* modify dockerfile to build within image

* Make active client balancing a configurable option

* Make active client balancing a configurable option

* With TLS throttle accepts if server is under heavy load - do not change non TLS behavior

* With TLS throttle accepts if server is under heavy load - do not change non TLS behavior

* Only run the tls-name-validation test if --tls is passed into runtest

* Only run the tls-name-validation test if --tls is passed into runtest

* Fix KeyDB not building with TLS < 1.1.1

* Fix KeyDB not building with TLS < 1.1.1

* update changelog to use replica as terminology

* update changelog to use replica as terminology

* update copyright

* update copyright

* update deb copyright

* update deb copyright

* call aeThreadOnline() earlier

* call aeThreadOnline() earlier

* Removed mergeReplicationId

* Removed mergeReplicationId

* acceptTLS is threadsafe like the non TLS version

* acceptTLS is threadsafe like the non TLS version

* setup Machamp ci

* setup Machamp ci

* make build_test.sh executable

* make build_test.sh executable

* PSYNC production fixes

* PSYNC production fixes

* fix the Machamp build

* fix the Machamp build

* break into tests into steps

* break into tests into steps

* Added multimaster test

* Added multimaster test

* Update ci.yml

Change min tested version to 18.04

* Update ci.yml

Change min tested version to 18.04

* fork lock for all threads, use fastlock for readwritelock

* fork lock for all threads, use fastlock for readwritelock

* hide forklock object in ae

* hide forklock object in ae

* only need to include readwritelock in ae

* only need to include readwritelock in ae

* time thread lock uses fastlock instead of std::mutex

* time thread lock uses fastlock instead of std::mutex

* set thread as offline when waiting for time thread lock

* set thread as offline when waiting for time thread lock

* update README resource links

* update README resource links

* Fix MALLOC=memkind build issues

* Fix MALLOC=memkind build issues

* Fix module test break

* Fix module test break

* Eliminate firewall dialogs on mac for regular and cluster tests.  There are still issues with the sentinel tests but attempting to bind only to localhost causes failures

* Eliminate firewall dialogs on mac for regular and cluster tests.  There are still issues with the sentinel tests but attempting to bind only to localhost causes failures

* remove unused var in networking.cpp

* remove unused var in networking.cpp

* check ziplist len to avoid crash on empty ziplist convert

* check ziplist len to avoid crash on empty ziplist convert

* remove nullptr subtraction

* remove nullptr subtraction

* cannot mod a pointer

* cannot mod a pointer

* need to include stdint for uintptr_t

* need to include stdint for uintptr_t

* use atomic_load for g_pserver->mstime

* use atomic_load for g_pserver->mstime

* Integrate readwritelock with Pro Code

* Integrate readwritelock with Pro Code

* Defensive asserts for RWLock

* Defensive asserts for RWLock

* Save and restore master info in rdb to allow active replica partial sync (#371)

* save replid for all masters in rdb

* expanded rdbSaveInfo to hold multiple master structs

* parse repl-masters from rdb

* recover replid info from rdb in active replica mode, attempt partial sync

* save offset from rdb into correct variable

* don't change replid based on master in active rep

* save and load psync info from correct fields

* Save and restore master info in rdb to allow active replica partial sync (#371)

* save replid for all masters in rdb

* expanded rdbSaveInfo to hold multiple master structs

* parse repl-masters from rdb

* recover replid info from rdb in active replica mode, attempt partial sync

* save offset from rdb into correct variable

* don't change replid based on master in active rep

* save and load psync info from correct fields

* placement new instead of memcpy

* placement new instead of memcpy

* Remove asserts, RW lock can go below zero in cases of aeAcquireLock

* Remove asserts, RW lock can go below zero in cases of aeAcquireLock

* Inclusive language

* Inclusive language

* call aeThreadOnline() earlier

* call aeThreadOnline() earlier

* Removed mergeReplicationId

* Removed mergeReplicationId

* Make active client balancing a configurable option

* Make active client balancing a configurable option

* With TLS throttle accepts if server is under heavy load - do not change non TLS behavior

* With TLS throttle accepts if server is under heavy load - do not change non TLS behavior

* acceptTLS is threadsafe like the non TLS version

* acceptTLS is threadsafe like the non TLS version

* PSYNC production fixes

* PSYNC production fixes

* Ensure we are responsive during storagecache clears

* Ensure we are responsive during storagecache clears

* Ensure recreated tables use the same settings as ones made at boot

* Ensure recreated tables use the same settings as ones made at boot

* Converted some existing PSYNC tests for multimaster

* Converted some existing PSYNC tests for multimaster

* Inclusive language fix

* Inclusive language fix

* Cleanup test suite

* Cleanup test suite

* Updated test replica configs so tests make sense

* Updated test replica configs so tests make sense

* active-rep test reliability

* active-rep test reliability

* Quick fix to make psync tests work

* Quick fix to make psync tests work

* Fix PSYNC test crashes

* Fix PSYNC test crashes

* Ensure we force moves not copies when ingesting bulk insert files

* Ensure we force moves not copies when ingesting bulk insert files

* Disable async for hget commands as it is not ready

* Disable FLASH

* Fix crash in save of masterinfo

* Fix musl/Alpine build failures

* Remove unnecessary libs

* update readme

* update readme

* remove Enterprise references

* Limit max overage to 20% during RDB save

* Delete COPYING to replace with BSD license

* update deb master changelog

* Update license

* Fix Readme typo from github org transition

Replace mention of scratch-file-path with db-s3-object

* Fix reference counting failure in the dict.  This is caused by std::swap also swapping refcounts

* Fix assertion in async rehash

* Prevent crash on shutdown by avoiding dtors (they are unnecessary anyways)

* Initialize noshrink, it was dangling

* Prevent us from starting a rehash when one wasn't already in progress.  This can cause severe issues for snapshots

* Avoid unnecessary rehashing when a rehash is abandoned

* Dictionary use correct acquire/release semantics

* Add fence barriers for the repl backlog (important for AARCH64 and other weak memory models)

* Silence TSAN errors on ustime and mstime.  Every CPU we support is atomic on aligned ints, but correctness matters

* Disable async commands by default

* Fix TSAN warnings on the repl backlog

* Merge OSS back into pro

* Fix unmerged files

* Fix O(n^2) algorithm in the GC cleanup logic

* Fix crash in expire when a snapshot is in flight.  Caused by a perf optimization getting the expire map out of sync with the val

* On Alpine we must have a reasonable stack size

* Revert ci.yml to unstable branch version

* Implements the soft shutdown feature to allow clients to cooperatively disconnect preventing disruption during shutdown

* Ensure clean shutdown with multiple threads

* update dockerfiles

* update deb pkg references and changelog

* update gem reference

* lpGetInteger returns int64_t, avoid overflow (#10068)

Fix #9410

Crucial for the ms and sequence deltas, but I changed all
calls, just in case (e.g. "flags")

Before this commit:
`ms_delta` and `seq_delta` could have overflown, causing `currid` to be wrong,
which in turn would cause `streamTrim` to trim the entire rax node (see new test)

* Fix issue #454 (BSD build break)

* Do not allow commands to run in background when in eval, Issue #452

* Fix certificate leak during connection when tls-allowlists are used

* Fix issue #480

* Fix crash running INFO command while a disk based backlog is set

* check tracking per db

* fix warnings

* Fix a race when undoConnectWithMaster changes mi->repl_transfer_s but the connection is not yet closed and the event handler runs

* Fix a race in processChanges/trackChanges with rdbLoadRio by acquiring the lock when trackChanges is set

* Fix ASAN use after free

* Additional fixes

* Fix integer overflow of the track changes counter

* Fix P99 latency issue for TLS where we leave work for the next event loop

tlsProcessPendingData() needs to be called before we execute queued commands because it may enqueue more commands

* Fix race removing key cache

* Prevent crash on load in long running KeyDB instances

* Fixes a crash where the server assertion failed when the key exists in DB during RDB load

* Remove old assertion which is commented out.

* avoid from instatiating EpochHolder multiple times to improve performance and cpu utilization

* avoid from instatiating EpochHolder multiple times to improve performance and cpu utilization

* src\redis-cli.c: fix potential null pointer dereference found by cppcheck

src\redis-cli.c:5488:35: warning: Either the condition
'!table' is redundant or there is possible null pointer dereference:
table. [nullPointerRedundantCheck]

* Fix Issue #486

* Workaround bug in snapshot sync - abort don't crash

* Improve reliability of async parts of the soft shutdown tests

* Improve reliability of fragmentation tests

* Verify that partial syncs do indeed occur

* Fix O(n) algorithm in INFO command

* Remove incorrect assert that fires when the repl backlog is used fully

* Make building flash optional

* Remove unneeded gitlab CI file

* [BUG] Moves key to another DB, the source key was removed if the move failed due to the key exists in the destination db #497 (#498)

Co-authored-by: Paul Chen <mingchen@Mings-MacBook-Pro.local>

* trigger repl_curr_off!= master_repl_offset assert failure when having pending write case

* use debug for logging the message instead

* rocksdb log using up the diskspace on flash (#519)

* Fix OpenSSL 3.0.x related issues. (#10291)

* Drop obsolete initialization calls.
* Use decoder API for DH parameters.
* Enable auto DH parameters if not explicitly used, which should be the
  preferred configuration going forward.

* remove unnecessary forward declaration

* remove internal ci stuff

* remove more internal ci/publishing

* submodule update step

* use with syntax instead

* bump ci ubuntu old ver as latest is now 22.04

* include submodules on all ci jobs

* install all deps for all ci jobs

Co-authored-by: Vivek Saini <vsaini@snapchat.com>
Co-authored-by: Christian Legge <christian@eqalpha.com>
Co-authored-by: benschermel <bschermel@snapchat.com>
Co-authored-by: John Sully <john@csquare.ca>
Co-authored-by: zliang <zliang@snapchat.com>
Co-authored-by: malavan <malavan@eqalpha.com>
Co-authored-by: John Sully <jsully@snapchat.com>
Co-authored-by: jfinity <38383673+jfinity@users.noreply.github.com>
Co-authored-by: benschermel <43507366+benschermel@users.noreply.github.com>
Co-authored-by: guybe7 <guy.benoish@redislabs.com>
Co-authored-by: Karthick Ariyaratnam (A) <k00809413@china.huawei.com>
Co-authored-by: root <paul.chen1@huawei.com>
Co-authored-by: Ilya Shipitsin <chipitsine@gmail.com>
Co-authored-by: Paul Chen <32553156+paulmchen@users.noreply.github.com>
Co-authored-by: Paul Chen <mingchen@Mings-MacBook-Pro.local>
Co-authored-by: Yossi Gottlieb <yossigo@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants