OAI-PMH data harvesting [KIWI]

Owned by Roman_Fedynyshyn
Last updated: Dec 24, 2021

Overview 

The purpose of these set of tests is to measure performance of Kiwi release. Find possible issues, bottlenecks. PERF-198 - Getting issue details... STATUS

Environment 

Software versions (Test 1-2)

  • mod-oai-pmh:3.7.0-SNAPSHOT.188.  
  • edge-oai-pmh:2.4.0
  • mod-source-record-manager:3.2.3
  • mod-source-record-storage:5.2.1
  • mod-inventory-storage:22.0.1
  • okapi:4.9.0

Original PTF dataset containing 1,212,039 underlying MARC records for 8.7M instances

Software versions (Test 3 with Bugfest Dataset)

  • mod-oai-pmh:3.6.1
  • edge-oai-pmh:2.4.0
  • mod-source-record-manager:3.2.6
  • mod-source-record-storage-5.2.5
  • mod-inventory-storage-22.0.3
  • okapi:4.9.0

Bugfest dataset containing 8,034,444 underlying MARC records for 8.3M instances

Summary

  • Kiwi release was able ho harvest 7,808,200 records in 19 hr 8 min (1M records per 2 hours and 15 min).
  • Average response time per request with resumption token 0.874s.
  • No memory or CPU issues were found (after the first couple of JIRAs below had been fixed)
  • KPIs: 
    • mod-oai-pmh CPU usage 120% (on data transferring) 100% on harvesting.
    • RDS CPU usage 80% on data transferring and ±15 % on harvesting
    • Memory usage 105-107% on mod-source-record-manager. 35% on mod-oai-pmh. No signs of memory leaks on related modules.
  • A few issues were found
    • OutOfMemory exception: MODOAIPMH-374 - Getting issue details... STATUS
    • Thread block issue: MODOAIPMH-374 - Getting issue details... STATUS
    • When instances didn't have underlying MARC records, multiple repeating calls from mod-edge-oai-pmh to mod-oai-pmh were occurred, resulting in the end-client receiving an timeout, see MODOAIPMH-383 - Getting issue details... STATUS  

Test flow

Test consist of few calls:

Initial call that was performed only once

/oai/records?verb=ListRecords&metadataPrefix=marc21_withholdings&apikey=[APIKey]

Subsequent harvesting calls:

/oai/records?verb=ListRecords&apikey=[APIKey]&resumptionToken=[resumptionToken]


These calls were performed repeatedly, harvesting 100 records each time until there is no more data in [tenant]_mod_oai_pmh.instances table to harvest.

[resumptionToken] was set to 100, returning in initial call response and in each harvesting call until there is no more records to harvest. When all data has being harvested - resumptionToken will not return with the response.


Issues detected during testing

1) OutOfMemory exception. fixed in scope of  MODOAIPMH-374 - Getting issue details... STATUS

2) Thread block issue. fixed in scope of  MODOAIPMH-374 - Getting issue details... STATUS

3) Client timeouts.   MODOAIPMH-383 - Getting issue details... STATUS   New issue appearing when we're starting DB transferring and harvesting process at the same time. It's leads to high load on DB and it responding with timeout

2021-12-01T10:02:42,566 ERROR [vert.x-eventloop-thread-0] MarcWithHoldingsRequestHelper Save instance Ids failed: Timeout.

io.vertx.core.impl.NoStackTraceThrowable: Timeout

Test results

Test 1

  • Total Underlying SRS records: 1,212,039
  • Duration: 4 hr 57 min
  • Records transferred: 4,770,043 (should be 8,415,303)
  • Calls performed 20,618 


We can see here unstable part of test. This spikes on chart showing extremely increased response times. which leads to throughput gaps. At this point we still not sure what was happening so we checked the logs of 

  • RDS response times: PGLogs.log
  • mod-oai-pmh
  • nginx-oai-pmh
  • edge-oai-pmh
  • okapi

At each point we have good response times and we can't see correlation between logs and this chart. 


Service CPU usage has reached ±200% while data transfer. And it's on 50-60% level during data harvesting. However during "unstable part" of test it has drop down to 20-25%.


  • Service memory usage is stable. There is no suspects for memory leak.


Notable observations:

  • While data transferring process is going on the background DB CPU usage has reached 70%-75%.
  • Data transferring process has failed in 10 minutes and transfer only 4770043 from 8M records.
  • Harvesting itself consumes 15% DB CPU.



Test 2

  • Total Underlying SRS records: 1,212,039
  • Duration: 4 hr 25 min
  • Records transferred: 3,815,867 (should be 8,415,303)
  • Calls performed - 22305

Results were the same as in test 1, showing consistency in failures due to missing a large number of underlying MARC records.








Test 3 (with Bugfest Dataset)

  • Underlying MARC records: 8,034,444
  • Records transferred -  8,213,392
  • Records harvested - 78082 X 100 = 7,808,200
  • Time spent : 19 hr 8 min

Average response time for call with resumption token 0.874 ms

With new data set there is no "unstable parts" in this test.  The results in this test is the best and accurate representation of OAI-PMH performance in Kiwi. 


CPU usage is stable and without big spikes. As you can see there is higher CPU usage at the beginning of the test. It's data transferring  process between mod-inventory-stirage and mod-oai-pmh.


We can see here that memory usage is stable and none of containers didn't fail. 





Notable observations:

  • OutOfMemory exception MODOAIPMH-374 and Thread block issue MODOAIPMH-374 were found and resolved early on in the testing cycle.
  • The remaining issue: unstable parts of first couple of tests were made by data set MODOAIPMH-383 
    • Instances didn't have underlying records and this causes multiple repeating calls from mod-edge-oai-pmh to mod-oai-pmh.
    • This leads to end client to wait until oai-pmh will find records with underlying records. And often client fail with 504 getaway timeout (load balancer timeout 400 seconds).
    • The workaround for this issue was by testing with a dataset that has underlying MARC records, which was Bugfest's. This is an edge case as most systems would have more MARC records than instances. It does not need to be resolved for the Kiwi release.