[IMPALA-1590] Remove workarounds for DCHECK in sorter.cc:549 (var_len_blocks[0]->Pin() fails to pin), when running parallel test

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: Impala 1.4.1, Impala 2.3.0
• Fix Version/s: Impala 2.7.0
• Component/s: Backend
• Labels: resource-management
• Target Version: Product Backlog
• Assignee: Tim Armstrong
• Reporter: Ippokratis Pandis
• Created: 09/Dec/14 23:19
• Updated: 13/Sep/16 00:43
• Resolved: 06/Jun/16 00:13

Description

I was running a parallel cancellation test and hit the DCHECK in L549 below.
I see that there are 7 var_len_blocks, but the Sorter::MinBuffersRequired is 6. It seems that (1) there were some large (varlen) strings which caused to reserve at least one additional var-len block; and (2) because of the parallel tests it failed to allocate the 7th block, which was not reserved. As a result we hit the DCHECK at L549.

A temporary fix is to not DCHECK but return OOM in cases where the number of var-len blocks is larger than the minimum reserved.

 527Status Sorter::Run::PrepareRead() {
...
 541  if (fixed_len_blocks_.size() > 0) {
 542    bool pinned;
 543    RETURN_IF_ERROR(fixed_len_blocks_[0]->Pin(&pinned));
 544    DCHECK(pinned);
 545  }
 546  if (has_var_len_slots_ && var_len_blocks_.size() > 0) {
 547    bool pinned;
 548    RETURN_IF_ERROR(var_len_blocks_[0]->Pin(&pinned));
 549=>  DCHECK(pinned);
 550  }

Issue Links

• is related to: IMPALA-1240 TestCancellationFullSort hits mem limit of 300MB with tpch_parquet
• is related to: IMPALA-2344 Remove workaround for sorter.cc:654: Check failed: pinned
• relates to: IMPALA-2369 Crash: impala::Sorter::Run::Init