[KUDU-716] Single-disk tpch_real_world with SF=500 should complete successfully

• Type: Sub-task
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: Private Beta
• Component/s: test
• Labels: roadmap-candidate
• Target Version/s: Public beta
• Assignee: Jean-Daniel Cryans
• Reporter: Todd Lipcon
• Created: 19/Apr/15 13:26
• Updated: 18/Jun/19 13:02
• Resolved: 15/Sep/15 12:11

Description

We now have a "tpch_real_world" benchmark that loads lineitem for an arbitrary scale factor and runs queries against it. In trying to run this on a single-disk system, we are running into a number of issues that prevent it from completing successfully:

• OOMs on the server side, due to:
  • poor default MRS size throttling behavior (per-MRS limits, no default setting)
  • live scanners retaining large amounts of memory from MRS (KUDU-18)
• OOMs on the client side (bugs like KUDU-715)
• timeouts during writes (eg due to client not respecting backpressure)
• timeouts during scans (unclear why - perhaps the server went into swap?)

We should make it a release criteria that tpch_real_world SF500 can complete even on underprovisioned hardware (single-HDD).

Issue Links

• relates to: KUDU-542 Make MRS throttling global
• relates to: KUDU-615 MRS/DMS hold on to anchors longer than necessary
• relates to: KUDU-720 Log GC never runs if MRS is under constant write pressure
• relates to: KUDU-715 tpch_real_world crash with bad_alloc in client
• relates to: KUDU-18 Handle "slow readers" not holding too much memory