In mongodb documentation with regards to "Unordered Bulk Write" we have "With an unordered list of operations, MongoDB can execute the operations in parallel, but this behavior is not guaranteed."
Using the java driver (v 3.2.0) it looks like
collection.bulkWrite(documents, new BulkWriteOptions().ordered(false) )
is not executed in parallel. What can trigger a parallel execution of bulkWrite ?
Related
I have multiple queries running on the same spark structured streaming session.
The queries are writing parquet records to Google Bucket and checkpoint to Google Bucket.
val query1 = df1
.select(col("key").cast("string"),from_json(col("value").cast("string"), schema, Map.empty[String, String]).as("data"))
.select("key","data.*")
.writeStream.format("parquet").option("path", path).outputMode("append")
.option("checkpointLocation", checkpoint_dir1)
.partitionBy("key")/*.trigger(Trigger.ProcessingTime("5 seconds"))*/
.queryName("query1").start()
val query2 = df2.select(col("key").cast("string"),from_json(col("value").cast("string"), schema, Map.empty[String, String]).as("data"))
.select("key","data.*")
.writeStream.format("parquet").option("path", path).outputMode("append")
.option("checkpointLocation", checkpoint_dir2)
.partitionBy("key")/*.trigger(Trigger.ProcessingTime("5 seconds"))*/
.queryName("query2").start()
Problem: Sometimes job fails with ava.lang.IllegalStateException: Race while writing batch 4
Logs:
Caused by: java.lang.IllegalStateException: Race while writing batch 4
at org.apache.spark.sql.execution.streaming.ManifestFileCommitProtocol.commitJob(ManifestFileCommitProtocol.scala:67)
at org.apache.spark.sql.execution.datasources.FileFormatWriter$.write(FileFormatWriter.scala:187)
... 20 more
20/07/24 19:40:15 INFO SparkContext: Invoking stop() from shutdown hook
This error is because there are two writers writing to the output path. The file streaming sink doesn't support multiple writers. It assumes there is only one writer writing to the path. Each query needs to use its own output directory.
Hence, in order to fix this, you can make each query use its own output directory. When reading back the data, you can load each output directory and union them.
You can also use a streaming sink that supports multiple concurrent writers, such as the Delta Lake library. It's also supported by Google Cloud: https://cloud.google.com/blog/products/data-analytics/getting-started-with-new-table-formats-on-dataproc . This link has instructions about how to use Delta Lake on Google Cloud. It doesn't mention the streaming case, but what you need to do is changing format("parquet") to format("delta") in your codes.
I'm trying to write the result of multiple operations into an AWS Aurora PostgreSQL cluster. All the calculations performs right but, when I try to write the result into the database I get the next error:
py4j.protocol.Py4JJavaError: An error occurred while calling o12179.jdbc.
: java.lang.StackOverflowError
at org.apache.spark.sql.catalyst.trees.TreeNode$$anonfun$2.apply(TreeNode.scala:256)
at org.apache.spark.sql.catalyst.trees.TreeNode$$anonfun$2.apply(TreeNode.scala:256)
at org.apache.spark.sql.catalyst.trees.CurrentOrigin$.withOrigin(TreeNode.scala:70)
at org.apache.spark.sql.catalyst.trees.TreeNode.transformDown(TreeNode.scala:255)
I already tried to increase cluster size (15 r4.2xlarge machines), change number of partitions for the data to 120 partitions, change executor and driver memory to 4Gb each and I'm facing the same results.
The current SparkSession configuration is the next:
spark = pyspark.sql.SparkSession\
.builder\
.appName("profile")\
.config("spark.sql.shuffle.partitions", 120)\
.config("spark.executor.memory", "4g").config("spark.driver.memory", "4g")\
.getOrCreate()
I don't know if is a Spark configuration problem or if it's a programming problem.
Finally I found the problem.
The problem was an iterative read from S3 creating a really big DAG. I changed the way I read CSV files from S3 with the following instruction.
df = spark.read\
.format('csv')\
.option('header', 'true')\
.option('delimiter', ';')\
.option('mode', 'DROPMALFORMED')\
.option('inferSchema', 'true')\
.load(list_paths)
Where list_paths is a precalculated list of paths to S3 objects.
I am running something along the lines of the following:
results = queries.map do |query|
begin
Neo4j::Session.query(query)
rescue Faraday::TimeoutError
nil
end
end
After a few iterations I get an unrescued Faraday::TimeoutError: too many connection resets (due to Net::ReadTimeout - Net::ReadTimeout) and Neo4j needs switching off and on again.
I believe this is because the queries themselves aren't aborted - i.e. the connection times out but Neo4j carries on trying to run my query. I actually want to time them out, so simply increasing the timeout window won't help me.
I've had a scout around and it looks like I can find my queries and abort them via the Neo4j API, which will be my next move.
Am I right in my diagnosis? If so, is there a recommended way of managing queries (and aborting them) from neo4jrb?
Rebecca is right about managing queries manually. Though if you want Neo4j to automatically stop queries within a certain time period, you can set this in your neo4j conf:
dbms.transaction.timeout=60s
You can find more info in the docs for that setting.
The Ruby gem is using Faraday to connect to Neo4j via HTTP and Faraday has a built-in timeout which is separate from the one in Neo4j. I would suggest setting the Neo4j timeout as a bit longer (5-10 seconds perhaps) than the one in Ruby (here are the docs for configuring the Faraday timeout). If they both have the same timeout, Neo4j might raise a timeout before Ruby, making for a less clear error.
Query management can be done through Cypher. You must be an admin user.
To list all queries, you can use CALL dbms.listQueries;.
To kill a query, you can use CALL dbms.killQuery('ID-OF-QUERY-TO-KILL');, where the ID is obtained from the list of queries.
The previous statements must be executed as a raw query; it does not matter whether you are using an OGM, as long as you can input queries manually. If there is no way to manually input queries, and there is no way of doing this in your framework, then you will have to access the database using some other method in order to execute the queries.
So thanks to Brian and Rebecca for useful tips about query management within Neo4j. Both of these point the way to viable solutions to my problem, and Brian's explicitly lays out steps for achieving one via Neo4jrb so I've marked it correct.
As both answers assume, the diagnosis I made IS correct - i.e. if you run a query from Neo4jrb and the HTTP connection times out, Neo4j will carry on executing the query and Neo4jrb will not issue any instruction for it to stop.
Neo4jrb does not provide a wrapper for any query management functionality, so simply setting a transaction timeout seems most sensible and probably what I'll adopt. Actually intercepting and killing queries is also possible, but this means running your query on one thread so that you can look up its queryId in another. This is the somewhat hacky solution I'm working with atm:
class QueryRunner
DEFAULT_TIMEOUT=70
def self.query(query, timeout_limit=DEFAULT_TIMEOUT)
new(query, timeout_limit).run
end
def initialize(query, timeout_limit)
#query = query
#timeout_limit = timeout_limit
end
def run
start_time = Time.now.to_i
Thread.new { #result = Neo4j::Session.query(#query) }
sleep 0.5
return #result if #result
id = if query_ref = Neo4j::Session.query("CALL dbms.listQueries;").to_a.find {|x| x.query == #query }
query_ref.queryId
end
while #result.nil?
if (Time.now.to_i - start_time) > #timeout_limit
puts "killing query #{id} due to timeout"
Neo4j::Session.query("CALL dbms.killQuery('#{id}');")
#result = []
else
sleep 1
end
end
#result
end
end
UPDATE I have put up a follow up question that contains updated scripts and and a clearer setup on neo4j performance compared to mysql (how can it be improved?). Please continue there./UPDATE
I have some problems verifying the performance claims made in the "graph databases" book (page 20) and in the neo4j (chapter 1).
To verify these claims I created a sample dataset of 100000 'person' entries with 50 'friends' each, and tried to query for e.g. friends 4 hops away. I used the very same dataset in mysql. With friends of friends over 4 hops mysql returns in 0.93 secs, while neo4j needs 65 -75 secs (on repeated calls).
How can I improve this miserable outcome, and verify the claims made in the books?
A bit more detail:
I run the whole setup on a i5-3570K with 16GB Ram, using ubuntu12.04 64bit, java version "1.7.0_25" and mysql 5.5.31, neo4j-community-2.0.0-M03 (I get a similar outcome with 1.9)
All code/sample data can be found on https://github.com/jhb/neo4j-experiements/ (to be used with 2.0.0). The resulting sample data in different formats can be found on https://github.com/jhb/neo4j-testdata.
To use the scripts you need a python with mysql-python, requests and simplejson installed.
the dataset is created with friendsdata.py and stored to friends.pickle
friends.pickle gets imported to neo4j using import_friends_neo4j.py
friends.pickle gets imported to mysql using import_friends_mysql.py
I add indexes on t_user_friend.* in mysql
I added "create index on :node(noscenda_name) in neo4j
To make life a bit easier the friends.*.bz2 contain sql and cypher statements to create those datasets in mysql and neo4j 2.0 M3.
Mysql performance
I first warm mysql up by querying:
select count(distinct name) from t_user;
select count(distinct name) from t_user;
Then, for the real meassurment I do
python query_friends_mysql.py 4 10
This creates the following sql statement (with changing t_user.names):
select
count(*)
from
t_user,
t_user_friend as uf1,
t_user_friend as uf2,
t_user_friend as uf3,
t_user_friend as uf4
where
t_user.name='person8601' and
t_user.id = uf1.user_1 and
uf1.user_2 = uf2.user_1 and
uf2.user_2 = uf3.user_1 and
uf3.user_2 = uf4.user_1;
and repeats this 4 hop query 10 times. The queries need around 0.95 secs each. Mysql is configured to use a key_buffer of 4G.
neo4j performance testing
I have modified neo4j.properties:
neostore.nodestore.db.mapped_memory=25M
neostore.relationshipstore.db.mapped_memory=250M
and the neo4j-wrapper.conf:
wrapper.java.initmemory=2048
wrapper.java.maxmemory=8192
To warm up neo4j I do
start n=node(*) return count(n.noscenda_name);
start r=relationship(*) return count(r);
Then I start using the transactional http endpoint (but I get the same results using the neo4j-shell).
Still warming up, I run
./bin/python query_friends_neo4j.py 3 10
This creates a query of the form (with varying person ids):
{"statement": "match n:node-[r*3..3]->m:node where n.noscenda_name={target} return count(r);", "parameters": {"target": "person3089"}
after the 7th call or so each call needs around 0.7-0.8 secs.
Now for the real thing (4 hops) I do
./bin/python query_friends_neo4j.py 4 10
creating
{"statement": "match n:node-[r*4..4]->m:node where n.noscenda_name={target} return count(r);", "parameters": {"target": "person3089"}
and each call takes between 65 and 75 secs.
Open questions/thoughts
I'd really like see the claims in the books to be reproducable and correct, and neo4j faster then mysql instead of magnitudes slower.
But I don't know what I am doing wrong... :-(
So, my big hopes are:
I didn't do the memory settings for neo4j correctly
The query I use for neo4j is completely wrong
Any suggestions to get neo4j up to speed are highly welcome.
Thanks a lot,
Joerg
2.0 has not been performance optimized at all, so you should use 1.9.2 for comparison.
(if you use 2.0 - did you create an index for n.noscenda_name)
You can check the query plan with profile start ....
With 1.9 please use a manual index or node_auto_index for noscenda_name.
Can you try these queries:
start n=node:node_auto_index(noscenda_name={target})
match n-->()-->()-->m
return count(*);
Fulltext indexes are also more expensive than exact indexes, so keep the exact auto-index for noscenda_name.
can't get your importer to run, it fails at some point, perhaps you can share the finished neo4j database
python importer.py
reading rels
reading nodes
delete old
Traceback (most recent call last):
File "importer.py", line 9, in <module>
g.query('match n-[r]->m delete r;')
File "/Users/mh/java/neo/neo4j-experiements/neo4jconnector.py", line 99, in query
return self.call(payload)
File "/Users/mh/java/neo/neo4j-experiements/neo4jconnector.py", line 71, in call
self.transactionurl = result.headers['location']
File "/Library/Python/2.7/site-packages/requests-1.2.3-py2.7.egg/requests/structures.py", line 77, in __getitem__
return self._store[key.lower()][1]
KeyError: 'location'
Just to add to what Michael said, in the book I believe the authors are referring to a comparison that was done in the Neo4j in Action book - it's described in the free first chapter of that book.
At the top of page 7 they explain that they were using the Traversal API rather than Cypher.
I think you'll struggle to get Cypher near that level of performance at the moment so if you want to do those types of queries you'll want to use the Traversal API directly and then perhaps wrap it in an unmanaged extension.
I am in need to hook a custom execution hook in Apache Hive. Please let me know if somebody know how to do it.
The current environment I am using is given below:
Hadoop : Cloudera version 4.1.2
Operating system : Centos
Thanks,
Arun
There are several types of hooks depending on at which stage you want to inject your custom code:
Driver run hooks (Pre/Post)
Semantic analyizer hooks (Pre/Post)
Execution hooks (Pre/Failure/Post)
Client statistics publisher
If you run a script the processing flow looks like as follows:
Driver.run() takes the command
HiveDriverRunHook.preDriverRun()
(HiveConf.ConfVars.HIVE_DRIVER_RUN_HOOKS)
Driver.compile() starts processing the command: creates the abstract syntax tree
AbstractSemanticAnalyzerHook.preAnalyze()
(HiveConf.ConfVars.SEMANTIC_ANALYZER_HOOK)
Semantic analysis
AbstractSemanticAnalyzerHook.postAnalyze()
(HiveConf.ConfVars.SEMANTIC_ANALYZER_HOOK)
Create and validate the query plan (physical plan)
Driver.execute() : ready to run the jobs
ExecuteWithHookContext.run()
(HiveConf.ConfVars.PREEXECHOOKS)
ExecDriver.execute() runs all the jobs
For each job at every HiveConf.ConfVars.HIVECOUNTERSPULLINTERVAL interval:
ClientStatsPublisher.run() is called to publish statistics
(HiveConf.ConfVars.CLIENTSTATSPUBLISHERS)
If a task fails: ExecuteWithHookContext.run()
(HiveConf.ConfVars.ONFAILUREHOOKS)
Finish all the tasks
ExecuteWithHookContext.run() (HiveConf.ConfVars.POSTEXECHOOKS)
Before returning the result HiveDriverRunHook.postDriverRun() ( HiveConf.ConfVars.HIVE_DRIVER_RUN_HOOKS)
Return the result.
For each of the hooks I indicated the interfaces you have to implement. In the brackets
there's the corresponding conf. prop. key you have to set in order to register the
class at the beginning of the script.
E.g: setting the PreExecution hook (9th stage of the workflow)
HiveConf.ConfVars.PREEXECHOOKS -> hive.exec.pre.hooks :
set hive.exec.pre.hooks=com.example.MyPreHook;
Unfortunately these features aren't really documented, but you can always look into the Driver class to see the evaluation order of the hooks.
Remark: I assumed here Hive 0.11.0, I don't think that the Cloudera distribution
differs (too much)
a good start --> http://dharmeshkakadia.github.io/hive-hook/
there are examples...
note: hive cli from console show the messages if you execute from hue, add a logger and you can see the results in hiveserver2 log role.