Investigate what happens if STL allocation causes OOM

[hermanlee]

Issue by maykov
Tuesday Apr 14, 2015 at 22:48 GMT
Originally opened as MySQLOnRocksDB#52

---

STL relies on exceptions to handle OOM. We should check that opreator new is not overloaded to return NULL. We should also add exception handlers into all high level handler functions.

[hermanlee]

Comment by spetrunia
Sunday May 24, 2015 at 07:54 GMT

---

I think I have accidentally hit this on a debug build. The following happened:

2015-05-24 00:36:35 10087 [Note] RocksDB: Finished filtering dropped index 202
2015-05-24 00:36:35 10087 [Note] RocksDB: Finished filtering dropped index 197
2015-05-24 00:36:35 10087 [Note] RocksDB: Finished filtering dropped index 198
terminate called after throwing an instance of 'std::bad_alloc'
  what():  std::bad_alloc
Aborted (core dumped)

[hermanlee]

Comment by mdcallag
Sunday May 24, 2015 at 22:11 GMT

---

I got that once

On Sunday, May 24, 2015, Sergei Petrunia notifications@github.com wrote:

I think I have accidentally hit this on a debug build. The following
happened:

2015-05-24 00:36:35 10087 [Note] RocksDB: Finished filtering dropped index 202
2015-05-24 00:36:35 10087 [Note] RocksDB: Finished filtering dropped index 197
2015-05-24 00:36:35 10087 [Note] RocksDB: Finished filtering dropped index 198
terminate called after throwing an instance of 'std::bad_alloc'
what(): std::bad_alloc
Aborted (core dumped)

—
Reply to this email directly or view it on GitHub
MySQLOnRocksDB#52 (comment)
.

Mark Callaghan
mdcallag@gmail.com

[hermanlee]

Comment by igorcanadi
Tuesday May 26, 2015 at 18:34 GMT

---

huh yeah. we never really looked at what happens when OOM. we should never corrupt the data, but rocksdb will likely crash hard.

[hermanlee]

Comment by maykov
Friday Jun 05, 2015 at 22:40 GMT

---

So it is important

[gunnarku]

Internal investigation completed. Behavior observed, documented, and understood for various scenarios. To decide if we'll change the status quo (SIGABRT because of an unhandled exception which leads to process teardown) will require data from the field and further discussion.