Reuse a byte buffer for holding XML #3118
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
andrewkroh
force-pushed
the
bugfix/wlb/reuse-buffer-for-xml
branch
from
4d14dde to
b65395c
Compare
|
I should probably re-run all these benchmarks at the same time. They were taken at different times using vagrant (some while on battery power). The amount of bytes being allocated is smaller, but the batch processing times are a getting slower. Overall they should be faster given the smaller allocs and shorter GC pauses. Almost all of the time spent in reading event logs is in the
|
|
jenkins, retest it |
1 similar comment
|
jenkins, retest it |
andrewkroh
force-pushed
the
bugfix/wlb/reuse-buffer-for-xml
branch
from
b65395c to
e5d1f60
Compare
|
Don't merge this yet. After repeatedly running the tests I can confirm that this PR is slower than what's in master. I generated a CPU profile while running the benchmark and noticed that the I wrote a really simple |
andrewkroh
added
in progress
Previously the data was read into a []byte encoded as UTF16. Then that
data was converted to []uint16 so that we can use utf16.Decode(). Then
the []rune slice was converted to a string which did another data copy.
The XML was unmarshalled from the string.
This PR changes the code to convert the UTF16 []byte directly to UTF8 and
puts the result into a reusable bytes.Buffer. The XML is then unmarshalled
directly from the data in buffer.
```
BenchmarkUTF16ToUTF8-4 2000000 1044 ns/op 4 B/op 1 allocs/op
```
Original
```
--- PASS: TestBenchmarkBatchReadSize (68.04s)
bench_test.go100: batch_size=10, total_events=20000, batch_time=5.682627ms, events_per_sec=1759.7494961397256, bytes_alloced_per_event=44 kB, total_allocs=4923840
bench_test.go100: batch_size=100, total_events=30000, batch_time=53.850879ms, events_per_sec=1856.9799018508127, bytes_alloced_per_event=44 kB, total_allocs=7354285
bench_test.go100: batch_size=500, total_events=25000, batch_time=271.118774ms, events_per_sec=1844.2101689350366, bytes_alloced_per_event=43 kB, total_allocs=6125665
bench_test.go100: batch_size=1000, total_events=30000, batch_time=558.03918ms, events_per_sec=1791.9888707455987, bytes_alloced_per_event=43 kB, total_allocs=7350324
```
After elastic#3113
```
--- PASS: TestBenchmarkBatchReadSize (71.85s)
bench_test.go:100: batch_size=10, total_events=30000, batch_time=5.713873ms, events_per_sec=1750.1264028794478, bytes_alloced_per_event=25 kB, total_allocs=7385820
bench_test.go:100: batch_size=100, total_events=30000, batch_time=52.454484ms, events_per_sec=1906.4147118480853, bytes_alloced_per_event=24 kB, total_allocs=7354318
bench_test.go:100: batch_size=500, total_events=25000, batch_time=260.56659ms, events_per_sec=1918.8952812407758, bytes_alloced_per_event=24 kB, total_allocs=6125688
bench_test.go:100: batch_size=1000, total_events=30000, batch_time=530.468816ms, events_per_sec=1885.124949550286, bytes_alloced_per_event=24 kB, total_allocs=7350360
```
After this PR
```
--- PASS: TestBenchmarkBatchReadSize (75.71s)
bench_test.go:100: batch_size=10, total_events=20000, batch_time=5.784644ms, events_per_sec=1728.7148526339736, bytes_alloced_per_event=14 kB, total_allocs=4863853
bench_test.go:100: batch_size=100, total_events=30000, batch_time=55.51756ms, events_per_sec=1801.2318985200359, bytes_alloced_per_event=14 kB, total_allocs=7264293
bench_test.go:100: batch_size=500, total_events=25000, batch_time=276.171282ms, events_per_sec=1810.4706484289702, bytes_alloced_per_event=14 kB, total_allocs=6050671
bench_test.go:100: batch_size=1000, total_events=30000, batch_time=619.758713ms, events_per_sec=1613.5311678950125, bytes_alloced_per_event=14 kB, total_allocs=7260333
```
Replace byte.Buffer usage with sys.ByteBuffer which is faster for the operations we need. ``` BenchmarkByteBuffer/byteBuffer-4 200000000 8.89 ns/op 0 B/op 0 allocs/op BenchmarkByteBuffer/bytes.Buffer-4 50000000 28.0 ns/op 0 B/op 0 allocs/op BenchmarkByteBufferGrow/byteBuffer-4 100000000 15.7 ns/op 2 B/op 0 allocs/op BenchmarkByteBufferGrow/bytes.Buffer-4 100000000 24.1 ns/op 2 B/op 0 allocs/op ```
andrewkroh
force-pushed
the
bugfix/wlb/reuse-buffer-for-xml
branch
from
e5d1f60 to
b0f98f6
Compare
|
I re-ran all the benchmarks after modifying this PR. The speed is about the same but now there will be less bytes allocated. |
andrewkroh
added
review
and removed
in progress
ruflin
approved these changes
Dec 7, 2016
suraj-soni
pushed a commit
to suraj-soni/beats
that referenced
this pull request
Dec 15, 2016
Previously the data was read into a []byte encoded as UTF16. Then that data was converted to []uint16 so that we can use utf16.Decode(). Then the []rune slice was converted to a string which did another data copy. The XML was unmarshalled from the string. This PR changes the code to convert the UTF16 []byte directly to UTF8 and puts the result into a reusable bytes.Buffer. The XML is then unmarshalled directly from the data in buffer. ``` BenchmarkUTF16ToUTF8-4 2000000 1044 ns/op 4 B/op 1 allocs/op ``` ``` git checkout 6ba7700 PS > go test github.com/elastic/beats/winlogbeat/eventlog -run TestBenc -benchtest -benchtime 10s -v === RUN TestBenchmarkBatchReadSize --- PASS: TestBenchmarkBatchReadSize (67.89s) bench_test.go:100: batch_size=10, total_events=30000, batch_time=5.119626ms, events_per_sec=1953.2676801000696, bytes_alloced_per_event=44 kB, total_allocs=7385952 bench_test.go:100: batch_size=100, total_events=30000, batch_time=51.366271ms, events_per_sec=1946.802795943665, bytes_alloced_per_event=44 kB, total_allocs=7354448 bench_test.go:100: batch_size=500, total_events=25000, batch_time=250.974356ms, events_per_sec=1992.2354138842775, bytes_alloced_per_event=43 kB, total_allocs=6125812 bench_test.go:100: batch_size=1000, total_events=30000, batch_time=514.796113ms, events_per_sec=1942.5166094834128, bytes_alloced_per_event=43 kB, total_allocs=7350550 PASS ok github.com/elastic/beats/winlogbeat/eventlog 67.950s git checkout 833a806 (elastic#3113) PS > go test github.com/elastic/beats/winlogbeat/eventlog -run TestBenc -benchtest -benchtime 10s -v === RUN TestBenchmarkBatchReadSize --- PASS: TestBenchmarkBatchReadSize (65.69s) bench_test.go:100: batch_size=10, total_events=30000, batch_time=4.858277ms, events_per_sec=2058.3429063431336, bytes_alloced_per_event=25 kB, total_allocs=7385847 bench_test.go:100: batch_size=100, total_events=30000, batch_time=51.612952ms, events_per_sec=1937.49816906423, bytes_alloced_per_event=24 kB, total_allocs=7354362 bench_test.go:100: batch_size=500, total_events=25000, batch_time=241.713826ms, events_per_sec=2068.561853801445, bytes_alloced_per_event=24 kB, total_allocs=6125757 bench_test.go:100: batch_size=1000, total_events=30000, batch_time=494.961643ms, events_per_sec=2020.3585755431961, bytes_alloced_per_event=24 kB, total_allocs=7350474 PASS ok github.com/elastic/beats/winlogbeat/eventlog 65.747s This PR (elastic#3118) PS > go test github.com/elastic/beats/winlogbeat/eventlog -run TestBenc -benchtest -benchtime 10s -v === RUN TestBenchmarkBatchReadSize --- PASS: TestBenchmarkBatchReadSize (65.80s) bench_test.go:100: batch_size=10, total_events=30000, batch_time=4.925281ms, events_per_sec=2030.341009985014, bytes_alloced_per_event=14 kB, total_allocs=7295817 bench_test.go:100: batch_size=100, total_events=30000, batch_time=48.976134ms, events_per_sec=2041.8108134055658, bytes_alloced_per_event=14 kB, total_allocs=7264329 bench_test.go:100: batch_size=500, total_events=25000, batch_time=250.314316ms, events_per_sec=1997.4886294557757, bytes_alloced_per_event=14 kB, total_allocs=6050719 bench_test.go:100: batch_size=1000, total_events=30000, batch_time=499.861923ms, events_per_sec=2000.5524605641945, bytes_alloced_per_event=14 kB, total_allocs=7260400 PASS ok github.com/elastic/beats/winlogbeat/eventlog 65.856s ```
monicasarbu
pushed a commit
that referenced
this pull request
Dec 19, 2016
* Rewrite elasticsearch connection URL (#3058) * Fix metricbeat service times-out at startup (#3056) * remove init collecting of processes * add changelog entry * Clarify that json.message_key is optional in Filebeat (#3055) I reordered the options based on importance (I put the optional config setting at the end). And I changed the wording to further clarify that the `json.message_key` setting is optional. Fixes #2864 * Document add_cloud_metadata processor (#3054) Fixes #2791 * Remove process.GetProcStatsEvents as not needed anymore (#3066) * Fix testing for 2x releases (#3057) * Update docker files to the last major with the most recent minor and bugfix version * Renamed files to Dockerfile-2x to not have to be renamed every time a new bugfix is released * Remove scripts and config files which are not needed anymore To run testsuite for 2x releases, run: `TESTING_ENVIRONMENT=2x make testsuite` * Remove old release notes files from packetbeat docs (#3067) * Update go-ucfg (#3045) - Update go-ucfg - add support for parsing lists/dictionaries from environment variables and via `-E` flag * Parse elasticsearch URL before logging it (#3075) * Fix the total CPU time in the Docker dashboard (#3085) (#3086) Part of #2629. The name of the field was changed, but not in the dashboard. (cherry picked from commit e271d9f) * Switch partition metricset from client to broker (#3029) Update kafka broker query - Switch paritition metricset from client to broker - on connect try to find the broker id (address must match advertised host). - check broker is leader before querying offsets - query offsets for all replicas - remove 'isr' from event, and replace with boolean flag `insync_replica` - replace `replicas` from event with per event `replica`-id - update sarama to get offset per replica id * Make error fields optional in partition event (#3089) * Update data.json * Make it clear in the docs that publish_async is still experimental (#3096) Remove example for publish_async from the docs * Remove metadata prefix from config as not needed (#3095) * Remove left over string in template test (#3102) * Fix typo in Dockerfile comment (#3105) * Document batch_read_size is experimental in Winlogbeat * Add benchmark test for batch_read_size in Winlogbeat (#3107) * Fix ES 2.x integration test (#3115) There was a test that was loading a mock template, and this template was assuming 5.x. * Pass `--always-copy` to virtualenv (#3082) virtualenv creates symlinks so `make setup` fails when ran on a network mounted fs. `--always-copy` copies files to the destination dir rather than symlinking. * Add project prefix for composer environment (#3116) This prefix is need to run tests with different environments in parallel so one does not affect the other. Like this 2x and snapshot builds should be able to coexist * Reduce allocations in UTF16 conversion (#3113) When decoding a UTF16 string contained in a buffer larger than just the string, more space was allocated than required. ``` BenchmarkUTF16BytesToString/simple_string-4 2000000 846 ns/op 384 B/op 3 allocs/op BenchmarkUTF16BytesToString/larger_buffer-4 2000000 874 ns/op 384 B/op 3 allocs/op BenchmarkUTF16BytesToString_Original/simple_string-4 2000000 840 ns/op 384 B/op 3 allocs/op BenchmarkUTF16BytesToString_Original/larger_buffer-4 1000000 3055 ns/op 8720 B/op 3 allocs/op ``` ``` PS C:\Gopath\src\github.com\elastic\beats\winlogbeat> go test -v github.com/elastic/beats/winlogbeat/eventlog -run ^TestBenchmarkBatchReadSize$ -benchmem -benchtime 10s -benchtest === RUN TestBenchmarkBatchReadSize --- PASS: TestBenchmarkBatchReadSize (68.04s) bench_test.go:100: batch_size=10, total_events=20000, batch_time=5.682627ms, events_per_sec=1759.7494961397256, bytes_alloced_per_event=44 kB, total_allocs=4923840 bench_test.go:100: batch_size=100, total_events=30000, batch_time=53.850879ms, events_per_sec=1856.9799018508127, bytes_alloced_per_event=44 kB, total_allocs=7354285 bench_test.go:100: batch_size=500, total_events=25000, batch_time=271.118774ms, events_per_sec=1844.2101689350366, bytes_alloced_per_event=43 kB, total_allocs=6125665 bench_test.go:100: batch_size=1000, total_events=30000, batch_time=558.03918ms, events_per_sec=1791.9888707455987, bytes_alloced_per_event=43 kB, total_allocs=7350324 PASS ok github.com/elastic/beats/winlogbeat/eventlog 68.095s PS C:\Gopath\src\github.com\elastic\beats\winlogbeat> go test -v github.com/elastic/beats/winlogbeat/eventlog -run ^TestBenchmarkBatchReadSize$ -benchmem -benchtime 10s -benchtest === RUN TestBenchmarkBatchReadSize --- PASS: TestBenchmarkBatchReadSize (71.85s) bench_test.go:100: batch_size=10, total_events=30000, batch_time=5.713873ms, events_per_sec=1750.1264028794478, bytes_alloced_per_event=25 kB, total_allocs=7385820 bench_test.go:100: batch_size=100, total_events=30000, batch_time=52.454484ms, events_per_sec=1906.4147118480853, bytes_alloced_per_event=24 kB, total_allocs=7354318 bench_test.go:100: batch_size=500, total_events=25000, batch_time=260.56659ms, events_per_sec=1918.8952812407758, bytes_alloced_per_event=24 kB, total_allocs=6125688 bench_test.go:100: batch_size=1000, total_events=30000, batch_time=530.468816ms, events_per_sec=1885.124949550286, bytes_alloced_per_event=24 kB, total_allocs=7350360 PASS ok github.com/elastic/beats/winlogbeat/eventlog 71.908s ``` * Fix for errno 1734 when calling EvtNext (#3112) When reading a batch of large event log records the Windows function EvtNext returns errno 1734 (0x6C6) which is RPC_S_INVALID_BOUND ("The array bounds are invalid."). This seems to be a bug in Windows because there is no documentation about this behavior. This fix handles the error by resetting the event log subscription handle (so events are not lost) and then retries the EvtNext call with maxHandles/2. Fixes #3076 * Fetch container stats in parallel (#3127) Currently fetching container stats is very slow as each request takes up to 2 seconds. To improve the fetching time if lots of containers are around, this creates the rrequests in parallel. The main downside is that this opens lots of connections. This fix should only temporary until the bulk api is available: moby/moby#25361 * Fix heartbeat not accepting `mode` parameter (#3128) * Remove fixed container names as not needed (#3122) Add beat name to project namespace * This makes sure different beats environment do not affect each other for example when Kafka is used * It also allows to run the testsuites of all the beats in parallel Introduce `stop-environment` command to stop all containers * Add doc for decode_json_fields processor (#3110) * Add doc for decode_json_fields processor * Use changed param names * Add example of decode_json_fields processor * Fix intro language about processors * Adding AmazonBeat to community beats (#3125) I created a basic version of amazonbeat, which reads data from an amazon product periodically. This beat does not yet publish to elasticsearch. * Reuse a byte buffer for holding XML (#3118) Previously the data was read into a []byte encoded as UTF16. Then that data was converted to []uint16 so that we can use utf16.Decode(). Then the []rune slice was converted to a string which did another data copy. The XML was unmarshalled from the string. This PR changes the code to convert the UTF16 []byte directly to UTF8 and puts the result into a reusable bytes.Buffer. The XML is then unmarshalled directly from the data in buffer. ``` BenchmarkUTF16ToUTF8-4 2000000 1044 ns/op 4 B/op 1 allocs/op ``` ``` git checkout 6ba7700 PS > go test github.com/elastic/beats/winlogbeat/eventlog -run TestBenc -benchtest -benchtime 10s -v === RUN TestBenchmarkBatchReadSize --- PASS: TestBenchmarkBatchReadSize (67.89s) bench_test.go:100: batch_size=10, total_events=30000, batch_time=5.119626ms, events_per_sec=1953.2676801000696, bytes_alloced_per_event=44 kB, total_allocs=7385952 bench_test.go:100: batch_size=100, total_events=30000, batch_time=51.366271ms, events_per_sec=1946.802795943665, bytes_alloced_per_event=44 kB, total_allocs=7354448 bench_test.go:100: batch_size=500, total_events=25000, batch_time=250.974356ms, events_per_sec=1992.2354138842775, bytes_alloced_per_event=43 kB, total_allocs=6125812 bench_test.go:100: batch_size=1000, total_events=30000, batch_time=514.796113ms, events_per_sec=1942.5166094834128, bytes_alloced_per_event=43 kB, total_allocs=7350550 PASS ok github.com/elastic/beats/winlogbeat/eventlog 67.950s git checkout 833a806 (#3113) PS > go test github.com/elastic/beats/winlogbeat/eventlog -run TestBenc -benchtest -benchtime 10s -v === RUN TestBenchmarkBatchReadSize --- PASS: TestBenchmarkBatchReadSize (65.69s) bench_test.go:100: batch_size=10, total_events=30000, batch_time=4.858277ms, events_per_sec=2058.3429063431336, bytes_alloced_per_event=25 kB, total_allocs=7385847 bench_test.go:100: batch_size=100, total_events=30000, batch_time=51.612952ms, events_per_sec=1937.49816906423, bytes_alloced_per_event=24 kB, total_allocs=7354362 bench_test.go:100: batch_size=500, total_events=25000, batch_time=241.713826ms, events_per_sec=2068.561853801445, bytes_alloced_per_event=24 kB, total_allocs=6125757 bench_test.go:100: batch_size=1000, total_events=30000, batch_time=494.961643ms, events_per_sec=2020.3585755431961, bytes_alloced_per_event=24 kB, total_allocs=7350474 PASS ok github.com/elastic/beats/winlogbeat/eventlog 65.747s This PR (#3118) PS > go test github.com/elastic/beats/winlogbeat/eventlog -run TestBenc -benchtest -benchtime 10s -v === RUN TestBenchmarkBatchReadSize --- PASS: TestBenchmarkBatchReadSize (65.80s) bench_test.go:100: batch_size=10, total_events=30000, batch_time=4.925281ms, events_per_sec=2030.341009985014, bytes_alloced_per_event=14 kB, total_allocs=7295817 bench_test.go:100: batch_size=100, total_events=30000, batch_time=48.976134ms, events_per_sec=2041.8108134055658, bytes_alloced_per_event=14 kB, total_allocs=7264329 bench_test.go:100: batch_size=500, total_events=25000, batch_time=250.314316ms, events_per_sec=1997.4886294557757, bytes_alloced_per_event=14 kB, total_allocs=6050719 bench_test.go:100: batch_size=1000, total_events=30000, batch_time=499.861923ms, events_per_sec=2000.5524605641945, bytes_alloced_per_event=14 kB, total_allocs=7260400 PASS ok github.com/elastic/beats/winlogbeat/eventlog 65.856s ``` * Fix make package for community beats (#3094) gopkg.in needs to be copied from the vendor directory of libbeat in the vendor directory * Auto generate modules list (#3131) This is to ensure no modules are forgotten in the future * Remove duplicated enabled entry from redis config (#3132) * Remove --always-copy from virtualenv and make it a param (#3136) In #3082 `--always-copy` was introduced. This caused issue on build on some operating systems. This PR reverts the change but makes `VIRTUALENV_PARAMS` a variable which can be passed to the Makefile. This allows anyone to set `--always-copy` if needed. * Adjust script to generate fields of type geo_point (#3147) * Fix for broken dashboard dependency in Cassandra Dashboard (#3146) The Cassandra Dashboard was linking to the wrong Cassandra visualisation. Some left over with : in the names were still inside Closes #3140 * Fix quotes (#3142) * Fix a print statement to be python 3 compliant (#3144) * Remove -prerelease from the repo names (#3153) * Add mongobeat to list of community beats (#3156) Mongobeat discovers instances in a mongo cluster and can be configured to ship multiple document types - from the commands db.stats() and db.serverStatus() * Update to most recent latest builds (#3161) * Merge snapshot and latest build for Logstash into 1 docker file * Pass certificate options to import dashboards script (#3139) * Pass certificate options to import dashboards script -cert for client certificate -key for client certificate key -cacert for certificate authority * Add -insecure flag to import_dashboards (#3163) * Improve speed and stability of CI builds (#3162) Loading and creating docker images takes quite a bit of time on the travis builds. Especially calls like apt-get update and install take lots of time and bandwidth and fail from time to time, as a host is not available. Following actions were taken: * Fake Kibana container is now based on alpine * Redis stunnel container was also switched to alpine * Add enabled config for prospectors (#3157) The enabled config allows easily to enable and disable a specific prospector. This is consistent with metricbeat where each modules has an enabled config. By default enabled is set to true. * Prototype Filebeat modules implementation (#3158) Contains the Nginx module, including the fields.yml and several pipelines. * Add edits for docker module docs (#3176) * Restructure and edit processors content (#3160) * Cleaned up Changelog in master (#3181) Added the 5.1.0 and 5.1.1 sections, removed duplicates. * metricbeat: enhance kafka broker matching (#3129) - compare broker names to hostname - try to lookup metricbeat host machine fqdn and compare to broker name - compare all ips of local machine with resolved broker name ips * Filebeat MySQL module (#3171) * Contains slowlog and errors filesets * Test files for two mysql versions (5.5 and 5.7) * Add support for built-in variables (e.g. `builtin.hostname`) * Contains a sample Kibana dashboard Part of #3159. * Fix #3167 change ownership of files in build/ (#3168) Add a new Makefile rule: fix-permissions fix-permissions runs a docker container that changes the ownership of all files from root to the user that runs the Makefile * Updating documentation to add udplogbeat (#3190) * Packer customize package info (#3188) * packer: Enable overriding of vendor and license * packer: customize URL of documentation link * packer: location of readme.md.j2 folder can be specified with PACKER_TEMPLATES_DIR * Filebeat syslog module (#3191) * Basic parsing of syslog fields * Supports multiline messages if the lines after the first one start with a space. * Contains a simple Kibana dashboard * Deprecate filters option in metrictbeat (#3173) * Add support for multiple paths per fileset (#3195) We generally need more than one path per OS, because the logs location is not always the same. For example, depending on the linux distribution and how you installed it, MySQL can have it's error logs in a number of default "paths". The solution is to configure them all, which means that Filebeat might try to access unexisting folders. This also improves the python prototype to accept multiple modules and to accept namespaced parameters. E.g.: ./filebeat.py --modules=nginx,syslog -M nginx.access.paths=... * case insensitive hostname comparison in kafka broker matching (#3193) - re-use common.LocalIPAddrs in partition module for resolving IPs - add missing net.IPAddr type switch to common.LocalIPAddrs - update matching to extract addresses early on using strings.ToLower => ensure case insensitive matching by lowercasing * Adds a couchbase module for metricbeat (#3081) * Export cpu cores (#3192) * Fix: Request headers with split_cookies enabled (#3065) * Add 3140 to changelog (#3207) (#3208) (cherry picked from commit 0f4103f)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Previously the data was read into a []byte encoded as UTF16. Then that
data was converted to []uint16 so that we can use utf16.Decode(). Then
the []rune slice was converted to a string which did another data copy.
Then the XML was unmarshalled from the string.
This PR changes the code to convert the UTF16 []byte directly to UTF8 and
puts the result into a reusable bytes.Buffer. The XML is then unmarshalled
directly from the data in buffer.
Original
After #3113
After this PR