- 13 1月, 2018 40 次提交
-
-
由 Heikki Linnakangas 提交于
We had allowed it in GPDB, when we had all the fancy persistent table and filespace mechanisms to track the created-but-not-yet-committed tablespace creation. The upstream mechanism is more crude, and can't handle that gracefully. For example, if you abort the CREATE TABLESPACE transaction, the directory structure would be left behind. Or worse, when you DROP TABLESPACE, the underlying directory is deleted immediately, even if the transaction aborts. Hence, put back the upstream limitation. That's been acceptable for Postgres, and is surely acceptable for GPDB as well.
-
由 Heikki Linnakangas 提交于
Backport from PostgreSQL 9.0: commit db64a9d20eeb0d3a52b7fb520a88bc68b73b374b Author: Bruce Momjian <bruce@momjian.us> Date: Tue Jul 20 18:14:25 2010 +0000 Properly replay CREATE TABLESPACE during crash recovery by deleting directory/symlink before creation. Report from Tom Lane. Backpatch to 9.0.
-
由 Ashwin Agrawal 提交于
This change help `FtsTestConnection()` to correctly report segment state. `probeWalRepPublishUpdate()` updating fts_status if primary/mirror gets marked as down. Also, removed all other states and code setting the same as its not used for anything.
-
由 Heikki Linnakangas 提交于
This is a far cry from actually making the tests work, but let's see what the next error is after this is fixed.
-
由 Heikki Linnakangas 提交于
-
由 Heikki Linnakangas 提交于
-
由 Asim R P 提交于
As a followup to the removal of multipass logic in commit 813b817c: * Make sure that recovery.conf file does not exist after promotion. The file is renamed to recovery.done, like in upstream. * Update gp_segment_configuration after promotion only if we are a master's standby. * We can't actually access the catalog without some more initialization. Pull as much initialization from the old Pass4 as possible -- some of it is likely not needed, but this seems to get things working for now. Author: Asim R P <apraveen@pivotal.io> Author: Jacob Champion <pchampion@pivotal.io>
-
由 C.J. Jameson 提交于
Author: C.J. Jameson <cjameson@pivotal.io>
-
由 C.J. Jameson 提交于
Speculative removals, as these tests were halfway through being ported into Behave / unit tests. These tests are not actively run in a pipeline; follow-up stories are being highlighted for upcoming work Author: C.J. Jameson <cjameson@pivotal.io>
-
由 C.J. Jameson 提交于
- Continuation of previous general filespace removal commits Author: C.J. Jameson <cjameson@pivotal.io>
-
由 Heikki Linnakangas 提交于
-
由 Heikki Linnakangas 提交于
-
由 Heikki Linnakangas 提交于
-
由 Heikki Linnakangas 提交于
The server-side code to exclude the listed files was accidentally removed, with the removal of filespace-related changes to basebackup.c. Resurrect it.
-
由 Heikki Linnakangas 提交于
Harmless, but compiler warnings are not nice.
-
由 Daniel Gustafsson 提交于
-
由 Heikki Linnakangas 提交于
The code used to wait for sent_location == replay_location, but that had similar problems that we had with the similar code in gp_replica_check earlier. Firstly, there is no guarantee that a checkpoint had been sent to the mirror, by the time that we check for it. So even if sent_location == replay_location, it was possible that the checkpoint hadn't even been sent to the mirror. That was highly unlikely in practice, though. Secondly, there are some corner cases where the primary would send half of a WAL record, up to a page boundary, but not the rest of the record, until something nudges it. I think that happens if the WAL record is not part of a transaction, so that there is no commit record to force it to be flushed. In that case, replay_location would not catch up to sent_location, because sent_location would point in the middle of a WAL record, and the standby can't replay half of a WAL record. That situation would persist until there's some new WAL activity to force it to be flushed, which could take longer than the timeout used in the test (possibly forever). To fix, memorize the WAL posititions on each segment, after issuing the CHECKPOINT, and wait until replay_location catches up to that position. That more accurately reflects what we are trying to accomplish: we want to wait until the checkpoint has been replayed, not necessarily that the standby is completely in sync with the primary. (gp_replica_check needed more strict logic, it wants the standby to really catch up to the primary, not just up to the latest checkpoint.) This hopefully fixes the occasional failures of the missing_xlog test we've been seeing in the pipeline.
-
由 Daniel Gustafsson 提交于
Include the diff between the first mismatching byte in the log output as we have it, and it can aid debugging (it just now helped me). Can't hurt.
-
由 Daniel Gustafsson 提交于
Commit f4c05df21d0906482e38010b15bac8221f52ee0b temporarily removed sequences from gp_replica_check due to them causing failures. Re-add them again as the masking should be fixed now.
-
由 Daniel Gustafsson 提交于
SEQ_LOG_VALS are pre-logged to avoid logging every fetch from a sequence, so we need to mask out last_value and log_cnt for sequence pages before running consistency checks.
-
由 Daniel Gustafsson 提交于
Makes filenames in mismatch messages consistently quoted and includes the length diff for ease of debugging.
-
由 Daniel Gustafsson 提交于
There were windows where we could return with open filehandles, ensure that we always close up and free up resources on retries before exiting.
-
由 Heikki Linnakangas 提交于
I'm not sure if we're going to keep these tests in the end, but might as well fix this for now.
-
由 Heikki Linnakangas 提交于
-
由 Ashwin Agrawal 提交于
-
由 Ashwin Agrawal 提交于
-
由 Ashwin Agrawal 提交于
-
由 Ashwin Agrawal 提交于
-
由 Heikki Linnakangas 提交于
There are more in other tests, but I suspect those other tests are obsolete in whole, so I'm just patching this one that caused a failure in the concourse pipeline.
-
由 Heikki Linnakangas 提交于
-
由 Heikki Linnakangas 提交于
-
由 Heikki Linnakangas 提交于
-
由 Heikki Linnakangas 提交于
-
由 Heikki Linnakangas 提交于
-
由 Heikki Linnakangas 提交于
-
由 Jimmy Yih 提交于
The walreceiver test establishes a walrep connection to the master's walsender. With the recent change to start standby master by default on gpdemo cluster, the mock walreceiver will be unable to create a connection to the walsender since max_wal_senders is hardcoded to 1. Moving this to run only on mirrorless cluster makes more sense since the assumption is to have an available walrep connection.
-
由 Jimmy Yih 提交于
This reverts commit 3214d80fabf7c208e40b2ec4f7ec69c97a9e7dd6.
-
由 Jimmy Yih 提交于
This reverts commit f38f024d7d314ae8ac100d862307a228b212c70e.
-
由 C.J. Jameson 提交于
- use the new fact that datadirs are in the gp_segment_configuration - fix a few things with the gpperfmon behave tests (mostly for macOS) --> the change to mgmt_utils.py is to do the config file manipulation natively in python --> the change to the gp_bash_functions.sh is to use ASCII ' characters so that python string comparison is happier Author: C.J. Jameson <cjameson@pivotal.io>
-
由 Heikki Linnakangas 提交于
It just doesn't mean "filespace" anymore, just plain "datadir".
-