/hive-sre

Primary LanguageJavaApache License 2.0Apache-2.0

Hive SRE Tooling

This application has 3 sub-programs:

  • sre is used to find potential 'Hive' performance issues caused by small files and excessive partitions.
  • u3 is used to review 'Hive 1/2' environments for Hive3 upgrade planning.
  • perf is used to check the throughput of a JDBC connection.

Trouble-Shooting

Supported Metastore DB's

Sub-Program Database Version Tested Notes
u3 MySql 5.6 Limited Recommend upgrading 5.7. This is the lower MySql supported env for HDP
5.7 Yes
8.0 No Not supported by HDP
MariaDb 10.1 No, but should work as 10.2 does
10.2 Yes
Postgresql 9.6 No, but should work
10 Yes Field Tested, May still be a few rough edges
11 No, but should work at 10 does
Oracle 12 Yes Field Tested, May still be a few rough edges
sre MySql 5.6 Limited Recommend upgrading 5.7. This is the lower MySql supported env for HDP
5.7 Partly Some sre reports use CTE in the SQL, which isn't supported in this version. Those report will error, the other will run fine.
8.0 No Not supported by HDP
MariaDb 10.1 No, but should work as 10.2 does
10.2 Yes
Postgresql 9.6 No, but should work
10 Yes Field Tested, May still be a few rough edges
11 No, but should work at 10 does
Oracle 12 Yes Field Tested, May still be a few rough edges

Ensure you have the database appropriate driver in the ${HOME}/.hive-sre/aux_libs directory.

I've tried to match supported DB's for HDP 2.6.5 and 3.1.x as much as I could.

Hadoop CLI

Running hive-sre-cli on the command line is an alias to the hadoopcli application here.

It is an interactive HDFS Client Command Line tool.

Ping Performance Tool

Included in this application suite is a ping performance tool that you can use to measure cluster host letancy. It is a MapReduce application that uses a list of hosts and will ping those hosts from the MR Map Task and record the results to HDFS.

The MR program take a few options:

usage: hadoop jar <jar-file> com.cloudera.utils.mapreduce.MRPingTool -d <output-dir> -hl <hdfs_host_list_file> [-c <count>] [-m <num_of_mappers>]
 -c,--count <arg>        Ping Count (The number of iterations we'll run
                         ping).  Each ping request makes 5 pings.
                         So if this value is 3, we'll do 3 sets of 5 pings
                         (15 total). Default 5
 -d,--directory <arg>    Output Directory [REQUIRED]
 -h,--help               Help
 -hl,--host-list <arg>   The host list file on HDFS. A text file with a
                         FQHN (full qualified host name) per
                         line.[REQUIRED]
 -m,--mappers <arg>      Number of Mappers.  To get coverage, should be
                         more than the compute node count.  But no
                         guarantee of even distribution, so best to over
                         subscribe. Default 2

Tested OS's

CentOS 7.6 (ping)

The ping output is parsed by the application and other OS/versions may yield different output, which we've not (yet) setup parsing for.

Let's assume:

  • You've downloaded the Hive setup/ingest/eval scripts
  • This has been tested again CentOS 7.
  • The user running the MR job has rights in each compute node to run ping.
  • That ping is allowed on the network and not blocked between hosts.
  • The MR Job is run with the right user credentials and from an Edgenode configured for the target cluster.
    • We need the hadoop MR libs to submit the job.
  • The user running the job has ACL's that allow them to write to the EXTERNAL db location you create with ping_ddl.sql.
  • Build a 'host list' file with ALL the FQHN (fully qualified hostnames) in the cluster that you want to test.
    • This is a text file with a single host FQHN per line. Similar to /etc/hosts.
    • Place the file (we'll use the name host_list.txt) in your HDFS home directory.
  • We are using the standard EXTERNAL and MANAGED Warehouse locations on HDFS
  • PING_DB = ping_perf
  • BATCH_ID = 2022-10-22_01

Procedure

  1. Run DDL Scripts in Beeline 2. beeline -f ping_ddl.sql --hivevar PING_DB=ping_perf
  2. Run MR Job, using the expected Partition Directory of the raw table as the output. The -m parameter controls the number of 'mappers' created. The intent is to get a map task on every compute node in the cluster. To do that, with perfect distribution (unlikely), you need at least as many mappers as there are compute nodes. We recommend 2-3x that incase some nodes get doubled up. There's no way to ensure every host runs a task, hence the over subscription here. Use the eval report to determine where the gaps where (if any) and run again. 4. hadoop jar /usr/local/hive-sre/lib/hive-sre-shaded.jar com.cloudera.utils.mapreduce.MRPingTool -m 40 -c 3 -hl host_list.txt -d /warehouse/tablespace/external/hive/ping_perf.db/raw/batch_id=2022-10-21_01
  3. Run the Ingest script in beeline 6. beeline -f ping_ingest.sql --hivevar PING_DB=ping_perf --hivevar BATCH_ID=2022-10-21_01
  4. Run the eval scripts in beeline --hivevar PING_DB=ping_perf --hivevar BATCH_ID=2022-10-21_01

Get the Binary

USE THE PRE-BUILT BINARY!!! You won't have the necessary dependencies to build this from scratch without downloading and building the 'Hadoop Cli'.

Don't Build, Download the LATEST binary here!!!

Download the LATEST Binary

On the edgenode:

  • Remove previous install directory rm -rf hive-sre-install
  • Expand the tarball tar zxvf hive-sre-dist.tar.gz.

    This produces a child hive-sre-install directory.

  • Two options for installation:
    • As the root user (or sudo), run hive-sre-install/setup.sh. This will install the hive-sre packages in /usr/local/hive-sre and create symlinks for the executables in /usr/local/bin. At this point, hive-sre should be available to all user and in the default path.
    • As the local user, run hive-sre-install/setup.sh. This will install the hive-sre packages in $HOME/.hive-sre and create symlink in $HOME/bin. Ensure $HOME/bin is in the users path and run hive-sre.

DO NOT RUN hive-sre from the installation directory.

If you install both options, your environment PATH will determine which one is run. Make note of this because an upgrade may not be reachable.

This will create and install the hive-sre and hive-sre-cli applications to your path.

Try it out on a host with default configs (if kerberized, get a ticket first):

hive-sre-cli

OR

hive-sre

Configuring hive-sre

See the config docs for details.

Running

To ease the launch of the application below, configure these core environment variables.

hive-sre u3|sre -all|-hdp2|-hdp3|-cdh`

NOTE It is NOT necessary to use the -cfg option if your config is $HOME/.hive-sre/cfg/default.

Using Encrypted Password in the config 'yaml'

Generate the Encrypted Password

Use the -pkey <password-key> and -p <password-to-encrypt options of hive-sre

hive-sre u3 -pkey cloudera -p have-a-nice-day

Will generate:

...
Encrypted password: HD1eNF8NMFahA2smLM9c4g==

Copy this encrypted password and place it in your configuration file for the connection password. Repeat for the other passwords, if it's different, and paste it in the configuration as well.

Running with Encrypted Passwords

Using the same -pkey <password-key> you used to generate the encrypted password, we'll run hive-sre

hive-sre u3 -all|-hdp2|-hdp3|-cdh -pkey cloudera ...

When the -pkey option is specified WITHOUT the -p option (used previously), hive-sre will understand to decrypt the configuration passwords before connecting to the resources. If you receive jdbc connection exceptions, recheck the -pkey and encrypted password from before.

NOTE: The encrypted password process is shared by u3, sre, and perf. It's not necessary to use different configs or password keys.

Checking the 'encrypted' password with Key

If you're not sure the password is correct, copy the 'encrypted' password from the config file and run:

hive-sre u3 -pkey <password-key> -dp <encrypted_password>

For example:

# Encrypt
dstreev@e01 ~ $ hive-sre u3 -pkey cloudera -p have-a-nice-day
APP_DIR: /usr/local/hive-sre/bin
Running Host instance
Application JAVA_OPTS=-Djavax.net.ssl.trustStore=/home/dstreev/bin/certs/gateway-client-trust.jks -Djavax.net.ssl.trustStorePassword=changeit
PRG_ARGS= "u3" "-pkey" "cloudera" "-p" "have-a-nice-day"
openjdk version "1.8.0_272"
OpenJDK Runtime Environment (build 1.8.0_272-b10)
OpenJDK 64-Bit Server VM (build 25.272-b10, mixed mode)
Launching: u3
Using Config: /home/dstreev/.hive-sre/cfg/default.yaml
1:Encrypted Password: HD1eNF8NMFahA2smLM9c4g==

# Decrypt
dstreev@e01 ~ $ hive-sre u3 -pkey cloudera -dp HD1eNF8NMFahA2smLM9c4g==
APP_DIR: /usr/local/hive-sre/bin
Running Host instance
Application JAVA_OPTS=-Djavax.net.ssl.trustStore=/home/dstreev/bin/certs/gateway-client-trust.jks -Djavax.net.ssl.trustStorePassword=changeit
PRG_ARGS= "u3" "-pkey" "cloudera" "-dp" "HD1eNF8NMFahA2smLM9c4g=="
openjdk version "1.8.0_272"
OpenJDK Runtime Environment (build 1.8.0_272-b10)
OpenJDK 64-Bit Server VM (build 25.272-b10, mixed mode)
Launching: u3
Using Config: /home/dstreev/.hive-sre/cfg/default.yaml
2:Decrypted Password: have-a-nice-day

Output

The output is a set of files with actions and error (when encountered). The files maybe txt files or markdown. You may want to use a markdown viewer for easier viewing of those reports. The markdown viewer needs to support github markdown tables .

UI Details for sre and u3

Only active processes will show up in the UI. The UI will refresh every second and display the current details below.

There are several 'processes' that are defined in u3. Each process will run 1 or more 'sub-processes'. The counters lists in the UI are specific to the 'process' and 'sub-processes' in that section.

The number of concurrent processes is controlled by the parallelism variable in the configuration yaml defined above.

  1. hive-sre version information
  2. Metastore RDBMS Type
  3. Thread Status a,b,c - d,e,f,j
    • (a) Core Pool Size
    • (b) Largest Pool Size
    • (c) Max Pool Size
    • (d) Active Thread Count
    • (e) Completed Thread Tasks
    • (f) Remaining Thread Queue
    • (j) Total Task Count
  4. Procedure Name
  5. Procedure Counts [Total/Completed]
    • Total is the full count of all tasks for that process
    • Completed is the number of tasks this procedure has completed.
  6. Procedure Check - Child of Procedure
  7. Procedure Check Counts - errors/successes
  8. Velocity - Total Time in Seconds process has been running.
  9. Velocity - The average number of Tasks completed per second since job started.
    • Note that not all Tasks are equal. Task times vary based on the content of the cluster and area being inspected. And can largely be effected by Namenode performance.

UI Details

Tips

  • Sorting results for loc_scan..
sort -k 1 --field-separator="|" loc_scan_missing_dirs.md > loc_scan_missing_dirs_sorted.txt