Overview

The Data Import Task Force (DITF) implements a feature that splits large input MARC files into smaller ones, resulting in smaller jobs, so that the big files could be imported and be imported consistently.  This document contains 1. Test with 1, 2, and 3 tenants' concurrent jobs with configurations the results of performance tests on the feature and also an analysis the feature's performance with respect to the baseline tests.  The following Jiras were implemented. 

PERF-644 - Getting issue details... STATUS PERF-645 - Getting issue details... STATUS PERF-647 - Getting issue details... STATUS PERF-646 - Getting issue details... STATUS PERF-671 - Getting issue details... STATUS

Summary

• The file-splitting feature is stable and offers more robustness to Data Import jobs even with the current infrastructure configuration. If there were failures, it's easier now to find the exact failed records to take actions on them. 
  • No stuck jobs in all tests performed.
  • There were errors (see below) in some partial jobs, but they still completed so the entire job status is "Completed with error".
  • Both of kinds of imports, create and update MARC BIBs worked well with this file-splitting feature enabled and also disabled. 
• There is no performance degradations, jobs not getting slower, on single-tenant imports. On multi-tenants imports, performance is be a little better
• Duration for DI correlates with number of the records imported (100k records- 38 min, 250k - 1 hour 32 min, 500k - 3 hours 29 min).
• Multitenant DI could be performed successfully for up to 9 jobs in parallel. If jobs are big they will start one by one in order for each tenant but processed in parallel on 3 tenants. Small DI (1 record) could be finished faster not in order. 
• No memory leak is suspected for all of the modules.
• Average CPU usage for mod-inventory -was 144%, mod-di-converter-storage was about 107%, and for all other modules did not exceed 100 %. We can observe spikes in CPU usage of mod-data-import at the beginning of the Data Import jobs up to 260%.  Big improvement over the previous version (without file-splitting) for 500K imports where mod-di-converter-storage's CPU utilization was 462% and other modules were above 100% and up to 150%. 
• Approximately DB CPU usage is up to 95%.

Recommendations and Jiras

1. One record on one tenant could be discarded with error: io.netty.channel.StacklessClosedChannelException. MODDATAIMP-748 - Getting issue details... STATUS Reproduces in both cases with and without splitting feature enabled in at least 30% of test runs with 500k record files and multitenant testing.
2. During the new Data Import splitting feature testing, items for update were discarded with the error: io.vertx.core.impl.NoStackTraceThrowable: Cannot get actual Item by id: org.folio.inventory.exceptions.InternalServerErrorException: Access for user 'data-import-system-user' (f3486d35-f7f7-4a69-bcd0-d8e5a35cb292) requires permission: inventory-storage.items.item.get. Less than 1% of records could be discarded due to missing permission for  'data-import-system-user'. Permission was not added automatically during the service deployment. I added permission manually to the database and the error does not occur anymore. MODDATAIMP-930 - Getting issue details... STATUS
3. UI issue, when canceled or completed with error Job progress bar cannot be deleted from the screen. MODDATAIMP-929 - Getting issue details... STATUS
4. Usage:
   • Should not use less than 1000 for RECORDS_PER_SPLIT_FILE. The system is stable enough to ingest 1000 records consistently and smaller amounts will incur more overheads, resulting in longer jobs' durations.  CPU utilization for mod-di-converter-storage for 500 RECORDS_PER_SPLIT_FILE(RPSF) = 160%, for 1000RPSF =180%, for 5K RPSF =380% and for 10K RPSF =433%, so in the case of selecting configurations 5K or 10K we recommend to add more CPU to mod-di-converter-storage service.
   • When toggling the file-splitting feature, mod-source-record-storage, mod-source-record-manager's tasks need to be restarted.
   • Keep in mind about the Kafka broker's disk size (as bigger jobs - up to 500K - can be run now), consecutive jobs may use up the disk quickly because the messages' retention time currently is set at 8 hours. For example with 300GB disk size, consecutive jobs of 250K, 500K, 500K sizes will exhaust the disk. 
5. More CPU could be allocated to mod-inventory and mod-di-converter-storage

Results

 * - One record on one tenant could be discarded with error: io.netty.channel.StacklessClosedChannelException. MODDATAIMP-748 - Getting issue details... STATUS  Reproduces in both cases with and without splitting features in at least 30% of test runs with 500k record files and multitenant testing.

 ** -  up to 10 items were discarded with the error: io.vertx.core.impl.NoStackTraceThrowable: Cannot get actual Item by id: org.folio.inventory.exceptions.InternalServerErrorException: Access for user 'data-import-system-user' (f3486d35-f7f7-4a69-bcd0-d8e5a35cb292) requires permission: inventory-storage.items.item.get. Less than 1% of records could be discarded due to missing permission for  'data-import-system-user'. Permission was not added automatically during the service deployment. I added permission manually to the database and the error does not occur anymore. MODDATAIMP-930 - Getting issue details... STATUS

Test 1,2. 100k, 250K, 500k and Multitenant MARC BIB Create

Memory Utilization

This has memory utilization increasing due to previous modules restarting (everyday cluster shot down process) no memory leak is suspected for DI modules.

MARC BIB CREATE

Test#1 100k, 250k, 500k records DI

Test#2 Multitenant  DI (9 concurrent jobs)

Service CPU Utilization 

MARC BIB CREATE

Average CPU usage for mod-inventory -was 144%, mod-di-converter-storage was about 107%, and for all other modules did not exceed 100 %. We can observe spikes in CPU usage of mod-data-import at the beginning of the Data Import jobs up to 260%.

Test#1 500k records DI

Test#2 Multitenant

Instance CPU Utilization

Test#1 500k records DI

Test#2 Multitenant DI (9 concurrent jobs)

RDS CPU Utilization 

MARC BIB CREATE

Approximately DB CPU usage is up to 95%

Test#1  500k records DI

Test#2 Multitenant  DI (9 concurrent jobs)

Maximal DB CPU usage is about 95%

RDS Database Connections

MARC BIB CREATE
 For DI  job Create- Maximum 535 connections count.

Test#1  500k records DI

Test#2 Multitenant

Test 3 With CI/CO 20 users and DI 25k records on each of the 3 tenants Splitting Feature Enabled & 

Splitting Feature Disabled

 * - Same approach testing DI: 3 DI jobs total on 3 tenants without CI/CO. Start the second job after the first one reaches 30%, and start another job on a third tenant after the first job reaches 60% completion. DI file size: 25k

Response time graph

With CI/CO 20 users and DI 25k records on each of the 3 tenants Splitting Feature Disabled

ocp3-mod-data-import:12

Data Import Robustness Enhancement

*   T1 - "00:33:35.1" Error T2 - "01:23:36.144" T3 - "01:16:26.391" on the first tenant proccesing stoped wit error "io.vertx.core.impl.NoStackTraceThrowable: Connection is not active now, current status: CLOSED "

it caused the spike of CPU utilization on Kafka (tenant cluster) up to 94% 

Instance CPU Utilization 

Test 1. Test with 1, 2, and 3 tenants' concurrent jobs with configuration RECORDS_PER_SPLIT_FILE = 500, 2 runs for each test. The maximal CPU Utilization value is 38%. 

Test 2. Test with 1, 2, and 3 tenants' concurrent jobs with configuration RECORDS_PER_SPLIT_FILE = 10K, 2 runs for each test. The maximal CPU Utilization value is 37%. 

Memory Utilization

Test 1. Test with 1, 2, and 3 tenants' concurrent jobs with configuration RECORDS_PER_SPLIT_FILE = 500, 2 runs for each test.

Most of the modules were stable during the test, and no memory leak is suspected for DI modules, only 2 modules increased memory consumption usage after the beginning of the tests

Memory utilization rich maximal value for mod-source-record-storage-b 88%  and for mod-source-record-manager-b 85%.

Test 2. Test with 1, 2, and 3 tenants' concurrent jobs with configuration RECORDS_PER_SPLIT_FILE = 10K, 2 runs for each test.

Service CPU Utilization 

Test 1. Test with 1, 2, and 3 tenants' concurrent jobs with configuration RECORDS_PER_SPLIT_FILE = 500, 2 runs for each test.

Test 2. Test with 1, 2, and 3 tenants' concurrent jobs with configuration RECORDS_PER_SPLIT_FILE = 10K, 2 runs for each test.

CPU utilization of  mod-di-converter-storage-b

RDS CPU Utilization 

Test 1. Test with 1, 2, and 3 tenants' concurrent jobs with configuration RECORDS_PER_SPLIT_FILE = 500, 2 runs for each test. Maximal  CPU Utilization = 95%

Test 2. Test with 1, 2, and 3 tenants' concurrent jobs with configuration RECORDS_PER_SPLIT_FILE = 10K, 2 runs for each test. Maximal  CPU Utilization = 94%

RDS Database Connections

Test 1. Test with 1, 2, and 3 tenants' concurrent jobs with configuration RECORDS_PER_SPLIT_FILE = 500, 2 runs for each test.

Test 2. Test with 1, 2, and 3 tenants' concurrent jobs with configuration RECORDS_PER_SPLIT_FILE = 10K, 2 runs for each test. Maximal  CPU Utilization = 94%

Retesting DI file-splitting feature on Poppy release

Retest the DI feature to be sure that the new changes have not affected performance negatively.  Retest the DI file-splitting feature for the following scenarios: PERF-681 - Getting issue details... STATUS

Brief comparison summary

The duration of the import date has increased, in particular(diff= Poppy time processing - Orchid time processing ):

• 250K MARC BIB Create PTF - Create 2 ---> 44 minutes
• 250K MARC BIB UpdatePTF - Updates Success - 1 -→ 45 minutes
• Multitenant MARC Create (100k, 50k, and 1 record)PTF - Create 2 -→1 hour 35 minutes
  • Check-Out without DI ~ 200ms
  • Check-In without DI ~ 650ms
  • Check-Out with DI ~ 770ms
  • Check-in with DI ~ 330ms

Resource utilization:

• Service CPU utilization on Poppy is about the same as on the Orchid;
• Memory utilization on Poppy is about the same as on the Orchid;
• RDS CPU Utilization  during all tests and on both releases was about 96%;
• The number of connections to DB on both releases were about the same from 550(Test 1.1) to 1200(Test 1.4).

Test 1.  Single tenant: create and update 250K file 

Test 1.4 With CI/CO 20 users and DI 25k records on each of the 3 tenants 

Resource utilization during testing

Test 1.1. Data-import of 250K records file with "PTF - Create 2" job profile

Service CPU Utilization 

The shark spike of CPU at the beginning of test 1, We see similar behavior in all of the DI tests. СPU consumption was uniform during the test.

Memory Utilization

The memory consumption was not affected, the mod-source-records-manager service increased the memory usage from 45% to 60% during the test, but after the test, the memory started to return to the pre-test value.

RDS CPU Utilization  

Consumption of the database CPU was 97% throughout the test

RDS Database Connections

The average number of DB connections during the test was about 550.

Test 1.2. Data-import of 250K records file with "PTF - Update" job profile

Service CPU Utilization 

СPU consumption was stable during the test, except mod-inventory service at the beginning of the test the CPU usage was about 140% at the end of the test CPU value was about 200%.   

Memory Utilization

The memory was stable and without memory leaks.

RDS CPU Utilization 

Consumption of the database CPU was 97% throughout the test

RDS Database Connections

The average number of DB connections during the test was about 550.

Test 1.3. Multitenant MARC Create (100k, 50k, and 1 record)

Service CPU Utilization 

СPU consumption was stable during the test. However, in the last hour of the test, the services mod-inventory and mod-quick-mark increare the CPU utilization by 75%

Memory Utilization

The memory was stable and without memory leaks.

RDS CPU Utilization 

Consumption of the database CPU was 96% throughout the test

RDS Database Connections

The average number of DB connections during the test was about 800.

Test 1.4. Data-import of 250K records file with "PTF - Update" job profile

Service CPU Utilization 

Memory Utilization 

The memory was stable and without memory leaks.

RDS CPU Utilization 

Consumption of the database CPU was 96% throughout the test

RDS Database Connections

The average number of DB connections during the test changed from 400 to 1200.

CICO responce time graph

Retesting DI file-splitting feature on Poppy release with Refresh Token Rotation (RTR) and file-splitting feature

The goal of the tests was to investigate how the file-splitting feature caused Data-import on Poppy release and the impact of Refresh Token Rotation (RTR). The tests were performed on ocp3(Poppy), pcp1(Poppy) and ncp5(Orchid)  environments.

PERF-723 - Getting issue details... STATUS Refresh Token Rotation (RTR)

Brief comparison summary

• Refresh Token Rotation configuration does not affect the data import process in any way, whether creating or updating a profile.
• In the Poppy release 250,000 records of data import with PTF - Create-2 job profile failed, and 50,000 records of data import with PTF - Updates Success - 1 job profile also failed in all of the tests, except configuration when FSF=ture;
• Data import works slowly on Poppy compared to the Orchid
• As the number of records in the file for data import increases, the processing time also increases. Up to 25,000 records, the duration of the data import is approximately the same.
• In the Poppy release data-import with an enabled file-splitting feature works slower compared to data-import with a disabled file-splitting feature.
• Data import is performed approximately 5% faster when the file-splitting feature parameters are absent in the task definition configuration.

Test results

Column with "pcp1 FSF false without token" has testing results on the configuration similar to "ocp3 FSF false without RTR token".

Resource utilization during testing

Service CPU utilization during the Data-import process

The next data import jobs were carried out
1) 5k_bib_Create 2) 10k_bib_Create  3) 25k_bib_Create 4) 50k_bib_Create 5) 50k_bib_Create 6) 100k_bib_Create 7) 50k_bib_Create 8) 25k_bib_Create 9) 25k_bib_Update 10) 50k_bib_Update(stopped)
CPU utilization was stable during all jobs, but some spikes of data-import jobs were at the beginning of all tests.
 

Memory Utilization

Most of the modules were stable during the test, and no memory leak is suspected for DI modules, except mod-inventory-b which consumed about 92% of memory during all DI processes. 

RDS CPU Utilization 

Maximal  CPU Utilization = 95%

RDS Database Connections

The maximal number of DB connections during the tests was about 580.

Database load

Top SQL queries

Appendix

Infrastructure ocp3  with the "Bugfest" Dataset

Records count :

• tenant0_mod_source_record_storage.marc_records_lb = 9674629
• tenant2_mod_source_record_storage.marc_records_lb = 0
• tenant3_mod_source_record_storage.marc_records_lb = 0
• tenant0_mod_source_record_storage.raw_records_lb = 9604805
• tenant2_mod_source_record_storage.raw_records_lb = 0
• tenant3_mod_source_record_storage.raw_records_lb = 0
• tenant0_mod_source_record_storage.records_lb = 9674677
• tenant2_mod_source_record_storage.records_lb = 0
• tenant3_mod_source_record_storage.records_lb = 0
• tenant0_mod_source_record_storage.marc_indexers =  620042011
• tenant2_mod_source_record_storage.marc_indexers =  0
• tenant3_mod_source_record_storage.marc_indexers =  0
• tenant0_mod_source_record_storage.marc_indexers with field_no 010 = 3285833
• tenant2_mod_source_record_storage.marc_indexers with field_no 010 = 0
• tenant3_mod_source_record_storage.marc_indexers with field_no 010 = 0
• tenant0_mod_source_record_storage.marc_indexers with field_no 035 = 19241844
• tenant2_mod_source_record_storage.marc_indexers with field_no 035 = 0
• tenant3_mod_source_record_storage.marc_indexers with field_no 035 = 0
• tenant0_mod_inventory_storage.authority = 4
• tenant2_mod_inventory_storage.authority = 0
• tenant3_mod_inventory_storage.authority = 0
• tenant0_mod_inventory_storage.holdings_record = 9592559
• tenant2_mod_inventory_storage.holdings_record = 16
• tenant3_mod_inventory_storage.holdings_record = 16
• tenant0_mod_inventory_storage.instance = 9976519
• tenant2_mod_inventory_storage.instance = 32
• tenant3_mod_inventory_storage.instance = 32 
• tenant0_mod_inventory_storage.item = 10787893
• tenant2_mod_inventory_storage.item = 19
• tenant3_mod_inventory_storage.item = 19

PTF -environment ocp3 

• 10 m6i.2xlarge EC2 instances located in US East (N. Virginia)us-east-1

• 2 database  instances, one reader, and one writer

  Name	API Name	Memory GIB	vCPUs	max_connections
  R6G Extra Large	db.r6g.xlarge	32 GiB	4 vCPUs	2731

• MSK ptf-kakfa-3
  
  • 4 m5.2xlarge brokers in 2 zones

  • Apache Kafka version 2.8.0

  • EBS storage volume per broker 300 GiB

  • auto.create.topics.enable=true
  • log.retention.minutes=480
  • default.replication.factor=3
• Kafka topics partitioning: - 2 partitions for DI topics

Before Splitting Feature released

Service versions for Splitting Feature test

Service versions for retesting Splitting Feature test on Poppy release. 

Methodology/Approach

To set splitting feature: Detailed Release Notes for Data Import Splitting Feature

Test 1: Manually tested 100k, 250k, and 500k records files started one by one on one tenant only.

Test 2: Manually tested 100k+50k+1 record files DI started simultaneously on every 3 tenants (9 jobs total).

Test 3: Run CICO on one tenant, DI jobs 3 tenants, including the one that runs CICO. Start the second job after the first one reaches 30%, and start another job on a third tenant after the first job reaches 60% completion. CICO: 20 users, DI file size: 25k

Test 4. To define the optimal value for RECORDS_PER_SPLIT_FILE(500, 1K, 2K, 5K) data-import job with PTF-Create-2 profile were run for 25K for 1 tenant simultaneously, for 2 tenants and for 3 tenants.