To put it in bit of an Indian context, database is not your daughter-in-law that you can blame it for every performance issue that occurs in the environment. But it does happen. Most of the time it is the database that is blamed for all such issues. Many times, the issues are in some other layer like OS, network or storage.
Faced this issue recently at one of the customer sites where performance in one of the databases went down suddenly. It was a 2 node RAC on 126.96.36.199 running on Linux 7 using some kind of Hitachi SSD storage array. There were no changes as per DBA, application, OS and storage teams. But something must have changed somewhere. Otherwise why would performance degrade just like that. I & my colleague checked some details and found that something happened in the morning a day before. Starting from that point in time, the execution time for all the commonly run queries shot up. Generally speaking, when all the queries are doing bad and you are sure that nothing has been changed on the database side, the reasons could be outside the database. But being a DBA, it is not easy to prove that. We took AWRs from good and bad times and the wait events section looked like this:
Now there is something clearly and terribly wrong with the details in the second snippet and in the first look it appears to be an IO issue. Av Rd(ms) in the File IO Stats section of the AWR reports was also showing really bad numbers for most of the data files, which have been fine two days ago.
The conference calls continued and we were not reaching anywhere. Storage team as usual said that everything was fine and there were no issues. Finally the discussion moved to multipathing and the teams started checking in that direction. There were errors like this in /var/log/messages
multipathd: asm!.asm_ctl_vbg1: failed to get path uid
multipathd: asm!.asm_ctl_vbg6: failed to get path uid
multipathd: asm!.asm_ctl_vbg9: failed to get path uid
That meant there was a problem with one of the paths from the database nodes to storage. They disabled the bad path for both the DB nodes and voila ! IO performance was back on track. It was multipathing that needed to be fixed.
So it is always not the database. It is unfair to always blame the DBA !
To be honest, Fernando Simon has already documented all the steps needed in ZDLRA patching . So this post is more like a reference post for me and it points to the links on his blog. One thing he could change though are the post titles. He also agrees 😉
ZDLRA patching is broadly divided into two parts. First part is where you patch the RA library and Grid & DB homes. Second part includes compute node & storage cell image patch and patches for IB/RoCE switches. Second part is exactly similar to Exadata except that it is bit restricted in terms of image versions that you can use. Only the versions that are certified for ZDLRA can be used. Also the RA library version and the Exadata image version should be compatible with each other. So if you are planning to patch only one part; RA library or the image, make sure that both the components stay compatible. The MOS note that has all these details is 1927416.1. This note should be the first place to go when you are planning to patch a ZDLRA. The steps for upgrade/patch, image patching are given in MOS note 2028931.1. There is another note 2639262.1 that discusses some of the known issues that you may face while doing the patching. It is important to review all three notes before you plan to patch.
The RA library patching part can be considered of two different types. This is an important difference. Make sure that you follow the right set of commands. When you are jumping between major versions say going from 12.x to 19.x, it is called an upgrade and the commands are like racli upgrade appliance –step=1. Fernando talks about this in detail in this post.
On the other hand, when you are not jumping between versions; say going from 19.x to 19.x only, it is called patching and the commands are like racli patch appliance –step=1. Fernando has discussed this in detail in this post.
The Exadata bit (image & switches patching) of it is exactly the same as we do in Exadata. Fernando talks about this in this post.
The RA library patching bit is pretty much automated and works fine most of the time. If you hit an issue, you may find the solution/workaround documented in one of the MOS notes.
Faced this while running installer for setting up a 2 node RAC setup (version 19.8) on an Oracle SuperCluster. The error reported in the log is:
[FATAL] [INS-44000] Passwordless SSH connectivity is not setup from the local node node1 to the following nodes:
[INS-06006] Passwordless SSH connectivity not set up between the following node(s): [node2]
From the error it appears that the ssh is not setup between two nodes but actually that is not the case. Here the error message is bit misleading. It turned out to be an issue with scp with openssh version 8.x. Running the setup with -debug option gives the clue:
<protocol error: filename does not match request>
The reason is a new check introduced in openssh version 8.x. It is explained here, here and here. MOS note 2555697.1 also talks about it.
Workaround is to pass the -T option to scp to ignore the new checks. You can rename the scp binary to something like scp.original and create a new shell script there like this:
mv scp scp.original
/usr/bin/scp.original -T $*
chmod 555 scp
This time, the install should succeed. You can revert the changes back once the install is done.
Earlier versions of OEDA didn’t allow you to have mixed cells in the configuration i.e. High Capacity (HC) and Extreme Flash (EF). The way to deal with that configuration was that deploy the system with either HC or EF cells and then manually configure the remaining cells.
I am not sure when did it change but the newer versions allow you have mixed type of cells in a single OEDA configuration. Once you select the hardware, there is an additional option called Enable Additional Storage, where you can select the other type of cells. The minimum number of cells has to be three to use this option. Also the cells that are at the bottom of the rack physically should be selected as main storage and the other cells should be added as additional storage as that is how OEDA builds the configuration files.
Once this is selected, on the Diskgroups screen, select Diskgroup layout as custom and you can create multiple diskgroups and select cells for each diskgroup (as EF & HC cells can’t be part of the same diskgroup).
Once the configuration is generated, it can be deployed with OneCommand without any manual intervention. A small feature but makes life easier by getting rid of all the manual steps.
In part 1, we discussed few things that you should take care before implementation of a ZDLRA. In this post, we will discuss few more things that you should review before or at the time of implementation:
If you are getting two ZDLRAs (one each for primary and standby sites), there are two ways they can be deployed. One scenario is where all the primary databases (or the database that have no standby) backup to RA at the primary site and then the data is replicated from primary RA to RA at the standby site. This works well for the DBs that have no standby database. For the DBs where there is a standby database, there is a better architecture that can be deployed. In that scenario, primary databases backup to primary RA and the standby databases backup to standby RA. That saves you all the traffic over replication network. Oracle has published a whitepaper on how to do this configuration. Few of the instructions in this paper are a bit dated but it gives a good overall idea of how to do the implementation.
Keep an eye on the features supported for different DB versions. An interesting one is that real-time redo shipping from standby databases is supported on 12c+ databases only. It is not supported for 11g. There could be other similar things. MOS note 1995866.1 has these details.
Depending upon the ZDLRA software version being deployed, it may need a minimum version of EM and the ZDLRA plugin. MOS note 2542836.1 has these details.
Make sure after discovering the the primary and standby databases in EM, their primary-standby relationship is reflected.
Real-time redo sent to ZDLRA is compressed but the archive logs backup will be compressed only if you use compression in the RMAN command. It is always good to include backup archivelog command with daily incremental job to make sure that no archive log is missed.
Many of the environments have separate networks for backup traffic. Make sure the backup traffic to ZDLRA uses DB server’s backup network. If that is not the case, you may need to add an explicit route on DB server for ZDLRA client/VIP/scan IPs.
There are going to be different users that you will need to use: one OS user for deploying the EM agent, one DB user that will be used to run the backups. Depending upon your environment, it may oracle OS user, SYS DB user or could be some other named user created for this purpose.
In next few posts, we will discuss some of the issues I have faced while doing ZDLRA implementation for some customers.
PS: Fernando Simon has written some brilliant posts related to ZDLRA on his blog. I highly recommend to review all of them. Brilliant stuff.