I have a batch job running with Flink on EMR which enriches some data stored as CSV on AWS S3 and indexes the tuples with Elasticsearch.
For some reason, one of the hosts is getting a lot more work than others. I tried to address that by hash partitioning on the several fields of the tuples but this doesn't make any difference: one of the nodes still gets more than the others. See host 40705 in screenshot below.
I need to distribute the indexing across the various nodes as well as possible to optimize throughput.
I tried using rebalance() but the result is the same. Any clues?
EDIT
Overview screen
Related
we've been playing a bit with Flink. So far we've been using Spark and standard M/R on Hadoop 2.x / YARN.
Apart from the Flink execution model on YARN, that AFAIK is not dynamic like spark where executors dynamically take and release virtual-cores in YARN, the main point of the question is as follows.
Flink seems just amazing: for streaming API's, I'd only say that it's brilliant and over the top.
Batch API's: processing graphs are very powerful and are optimised and run in parallel in a unique way, leveraging cluster scalability much more than Spark and others, optiziming perfectly very complex DAG's that share common processing steps.
The only drawback I found, that I hope is just my misunderstanding and lack of knowledge is that it doesn't seem to prefer data-local processing when planning the batch jobs that use input on HDFS.
Unfortunately it's not a minor one because in 90% use cases you have a big-data partitioned storage on HDFS and usually you do something like:
read and filter (e.g. take only failures or successes)
aggregate, reduce, work with it
The first part, when done in simple M/R or spark, is always planned with the idiom of 'prefer local processing', so that data is processed by the same node that keeps the data-blocks, to be faster, to avoid data-transfer over the network.
In our tests with a cluster of 3 nodes, setup to specifically test this feature and behaviour, Flink seemed to perfectly cope with HDFS blocks, so e.g. if file was made up of 3 blocks, Flink was perfectly handling 3 input-splits and scheduling them in parallel.
But w/o the data-locality pattern.
Please share your opinion, I hope I just missed something or maybe it's already coming in a new version.
Thanks in advance to anyone taking the time to answer this.
Flink uses a different approach for local input split processing than Hadoop and Spark. Hadoop creates for each input split a Map task which is preferably scheduled to a node that hosts the data referred by the split.
In contrast, Flink uses a fixed number of data source tasks, i.e., the number of data source tasks depends on the configured parallelism of the operator and not on the number of input splits. These data source tasks are started on some node in the cluster and start requesting input splits from the master (JobManager). In case of input splits for files in an HDFS, the JobManager assigns the input splits with locality preference. So there is locality-aware reading from HDFS. However, if the number of parallel tasks is much lower than the number of HDFS nodes, many splits will be remotely read, because, source tasks remain on the node on which they were started and fetch one split after the other (local ones first, remote ones later). Also race-conditions may happen if your splits are very small as the first data source task might rapidly request and process all splits before the other source tasks do their first request.
IIRC, the number of local and remote input split assignments is written to the JobManager logfile and might also be displayed in the web dashboard. That might help to debug the issue further. In case you identify a problem that does not seem to match with what I explained above, it would be great if you could get in touch with the Flink community via the user mailing list to figure out what the problem is.
I have performance concern and want a suggestion that which will be best, Multi Core or Multi Instance(with different port)? Lets have a look on My Case First:
Currently I am running solr with multiple core and its running OK. There is only one issue that sometime it goes "out of heap memory while processing facets fields", then I have to restart the solr. ( To minimize the no. of restarts, I starts the solr with high memory : java -Xms1000M -Xmx8000M -jar start.jar )
I have amazon ec2 instance with 8core-2.8GHtz /15GB Ram with optimized hard disk.
I have many database-tables(about 100) and have to create different schemas for each(leads to create different core).
Each table have millions of documents, with 7-9 indexed fields and 10-50 stored fields in each document.
My web portals should handle very high traffic (currently I m having 10 request/second, may increase to 50-100/second). I know 'solr' can handle that but it is to just inform you that I am concern about every-smallest performance issue also
Searching solr by PHP and CURL in to specific core, so there is no problem in searching in different solr instance also.
Question:
As per as I know Solr handles one request at a time. So I think if I create multiple instance of solr and starts those at different port, then my web portal can handle more request at a time. (if user search in different table).
So, what you will suggest me? Multi Core in Single Solr Instance? or Multiple Instances with Single/Dual Core in each?
Is there any problem in having multiple solr instances running at different ports?
NOTE: Here, I can/may/will combine less-searched-core(s)/small-core(s) in one instance AND heavy-traffic-core(s) in separate instance OR two-three-heavy-traffic-core in one-instance etc. Coz, creating different Instances for each table(~100 here) will take too much hardware resources.
As I didn't got any answer since more then week AND I had also tried many case with solr (and also read some articles), I want to share my experience as answer to my own question. This may/will help to future viewer. I tried on serverfault also with no success.
Solr can handle more request at a time.
I have tested it by running a long query [qTime=7203, approx. 7sec] and several small-queries-after-long-one [qTime=30], solr respond for small-queries first even they ran after long-one.
This point gives much reason in answer: Use single solr instance with multiple core. Just assign High memory to JVM.
Other Points:
1. Each solr instance will require RAM, so running multiple instances will require more resources, which will be expensive. And if you are using facets, sort fields then you need to allocate more RAM to each instance.
As you can see in my case I need to start the solr with high memory(8GB). You can see a case for Danish Web Archive, Which uses multiple instances and allocated 9GB RAM to each and having 256GM total RAM.
2. You can run multiple instances of solr on different PORT by java -Djetty.port=8984 -jar start.jar. Everything running ok BUT I got one problem.
While indexing it may give "not enough memory error" and then solr instance will be killed. So you again need to start second instance with high memory, which will leads to more RAM requirement.
3. Solr Resource Requirement and Performance Problem can be understand here. According to this 64bit environment and 12GB RAM is recommended for good performance. Solr Optimization are explained here.
Solr 3x "Repeaters" and Multiple Data Centers:
Solr 3x let a node behave as both a slave and master, pull from one master, and then feed copies downstream to its own slaves. This was so common/useful it even had a name, a "Repeater".
This was useful if you wanted span multiple data centers. You could have the real master in data center A (DCA), and a "repeater" in data center B (DCB). That repeater would then grab content from DCA and feed all of the other nodes in DCB, saving on bandwidth.
Suppose you want to upgrade this setup to Solr 4x and SolrCloud. (Note that Solr 4x still supports Solr 3x-style legacy replication)
It's said that you should NOT have a single SolrCloud cluster span disparate data centers. So data center B should have it's own SolrCloud.
One idea is to have the DCA -> DCB link still use Solr 3x-style Master/Slave replication. And then the "repeater" in DCB, being also a SolrCloud node, would automatically be propagated to other nodes.
Main question:
Can a Solr node participate in both Solr 3x-style master/slave mode (as a slave) and also be part of a SolrCloud cluster? And if so, how is this configured?
Complications:
In the simple case, if it's just 1 shard with replicas, it's easy to see how that might work in terms of data. It's a little less clear if you have multiple shards in DCB, how do I tell each shard to only replicate its own share of data? Note that SolrCloud normally replicates via transactions, whereas 3x uses binary indices.
Another complexity is if you're doing replication. How do you tell just the master node for each shard to pull from the remote DCA node?
Alternatives:
On solution is to upgrade to 4x but continue using 3x-style replication in DCB, so just don't use SolrCloud.
I realize that another solution would be to have the data feed send it's updates to both data centers, or usE something like RabbitMQ. For the sake of this question, let's assume thats not an option (long story...)
Maybe there's some other way I haven't thought of?
Has anybody actually tried having SolrCloud span data centers? How horrible is it?
Somebody must have asked this question before!
But I've looked on Google and, although it finds tons of pages with the keywords, I haven't seen this specific "hybrid" mode fleshed out. I found one thread from 2013 but it didn't really talk about the configuration and complexity.
To answer your first question, a Solr slave in 3.X style cannot be a node in a Solr Cloud. The reason is the slave in a master/slave 3.X Solr config simply replicates, byte for byte, all the index files on the master. That's all it does. It can, in the repeater config, then also be a master for others to replicate from, or be a dedicated query slave or both. But that's it.
A node in a Solr Cloud config is a full participant in a distributed computing cluster where indexing is generally intended to be distributed across all nodes, and all nodes participate in queries. It's a very powerful feature which automatically handles failed nodes and significantly eases the work load of scaling up that was very manual in 3.X style.
However, part of what you pay for that is increased complexity (Zookeeper), requirements for lower latency inter-node communications (because all the nodes now talk to each other and to Zookeeper) and the loss of the simplicity of Master/Slave replication.
At 20M docs you are well within the constraints of a single node master index with an effectively unlimited number of slaves and therefor very high query capacity. I do this today with a production environment where each master has on the order of 60M docs in it with no significant problems.
The question is do you need NRT, multi-node indexing, automated failover, the ability to autoscale well past 100M docs? If so then Master/Slave it probably not going to work for you.
You could take a look at writing the same data to two different Solr Cloud clusters, one in each datacenter. You could do that directly, or use something like Apache Flume to do it for you - in either there are some issues with doing this and so the real question is are dealing with those issues worth it to get the added benefit of Solr Cloud?
My client does not have the budget to setup and maintain a SOLR server to use in their production environment. If I understand the Sitecore 7 Content Search API correctly, it is not a big deal to configure things to use Lucene instead. For the most part the configuration will be similar and the code will be the same, and a SOLR server can be swapped in later.
The site build has
faceted search page
listing components on landing and on other pages that will leverage the Content Search API
buckets with custom facets
The site has around 5,000 pages and components not including media library items. Are there any concerns about simply using Lucene?
The main question is, when, during your architecture or design phase do you know that you should definitely choose SOLR over Lucene? What are the major signs that lead you recommend that?
I think if you are dealing with a customer on a limited budget then Lucene will work perfectly well and perform excellently for the scale of things you are doing. All the things you mention are fully supported by the implementation in Lucene.
In a Sitecore scenario I would begin to consider Solr if:
You need to index a large number of items - id say 50 thousand upwards - Lucene is happy with these sorts of number but Solr has improved query caching and is designed for these large numbers of items.
The resilience of the search tier is of maximum business importance (ie the site is purely driven by search) - Solr provides a more robust replication/sharding and failover system with SolrCloud.
Re-purposing of the search tier in other application is important (non Sitecore) - Solr is a search application so can be accessed over HTTP with XML/JSON etc which makes integration with external systems easier.
You need some specific additional feature of Solr that Lucene doesn't have.
.. but as you say if you want swap out Lucene for Solr at a later phase, we have worked hard to make sure that the process as simple as possible. Worth noting a few points here:
While your LINQ queries will stay the same your configuration will be slightly different and will need attention to port across.
The understanding of how Solr works as an application and how the schema works is important to know but there are some great books and a wealth of knowledge out there.
Solr has slightly different (newer) analyzers and scoring mechanisms so your search results may be slightly different (sometimes customers can get alarmed by this :P)
.. but I think these are things you can build up to over time and assess with the customer. Im sure there are more points here and others can chime in if they think of them. Hope this helps :)
Stephen pretty much covered the question - but I just wanted to add another scenario. You need to take into account the server setup in your production environment. If you are going to be using multiple content delivery servers behind a load balancer I would consider Solr from the start, as trying to make sure that the Lucene index on each delivery server is synchronized 100% of the time can be painful.
I would recommend planning an escape plan from Lucene as early as you start thinking about multiple CDs and here is why:
A) Each server has to maintain its own index copy:
Any unexpected restart might cause a few documents not to be added to the index on the one box, making indexes different from server to server.
That would lead to same page showing differently by CDs
Each server must perform index updates - use CPU & disk space; response rate drops after publish operation is over =/
According to security guide, CDs should have Sitecore Shell UI removed, so index cannot be easily rebuilt from Control Panel =\
B) Lucene is not designed for large volumes of content. Each search operation does roughly following:
Create an array with size equal to total number of documents in the index
If document matches search, set flag in the array
While this works like a charm for low sized indexes (~10K elements), huge performance degradation is produced once the volume of content grows.
The allocated array ends in Large Object Heap that is not compacted by default, thereby gets fragmented fast.
Scenario:
Perform search for 100K documents -> huge array created in memory
Perform one more search in another thread -> one more huge array created
Update index -> now 100K + 10 documents
The first operation was completed; LOH has space for 100K array
Seach triggered again -> 100K+10 array is to be created; freed memory 'hole' is not large enough, so more RAM is requested.
w3wp.exe process keeps on consuming more and more RAM
This is the common case for Analytics Aggregation as an index is being populated by multiple threads at once.
You'll see a lot of RAM used after a while on the processing instance.
C) Last Lucene.NET release was done 5 years ago.
Whereas SOLR is actively being developed.
The sooner you'll make the switch to SOLR, the easier it would be.
I have the following scenario:
Around 70 million of equipments send a signal every 3~5 minutes to
the server sending its id, status (online or offiline), IP, location
(latitude and longitude), parent node and some other information.
The other information might not be in an standard format (so no schema for me) but I still need to query it.
The equipments might disappear for some time (or forever) not sending
signals in the process. So I need a way to "forget" the equipments if
they have not sent a signal in the last X days. Also new equipments
might come online at any time.
I need to query all this data. Like knowing how many equipments are offline on a specific region or over
an IP range. There won't be many queries running at the same time.
Some of the queries need to run fast (less than 3 min per query) and
at the same time as the database is updating. So I need indexes on
the main attributes (id, status, IP, location and parent node). The
query results do not need to be 100% accurate, eventual consistency
is fine as long as it doesn't take too long (more than 20 min on
avarage) for them to appear in the queries results.
I don't need
persistence at all, if the power goes out it's okay to lose
everything.
Given all this I thought of using a noSQL approach maybe MongoDB or CouchDB since I have experience with MapReduce and Javascript but I don't know which one is better for my problem (I'm gravitating towards CouchDB) or if they are fit at all to handle this massive workload. I don't even know if I actually need a "traditional" database since I don't need persistence to disk (maybe a main-memory approach would be better?), but I do need a way to build custom queries easily.
The main problem I detect are the following:
Need to insert/update lots of tuples really fast and I don't know
beforehand if the signal I receive is already in the database or not.
Almost all of the signals will be in the same state as they were the
last time, so maybe query by id and check to see if the tuple changed if not do nothing, if it did update?
Forgeting offline equipments. A batch job that runs during the night
removing expired tuples would solve this problem.
There won't be many queries running at the same time, but they need
to run fast. So I guess I need to have a cluster that perform a
single query on multiple nodes of the cluster (does CouchDB MapReduce
splits the workload to multiple nodes of the cluster?). I'm not
enterily sure I need a cluster though, could a single more expensive
machine handle all the load?
I have never used a noSQL system before, but I have theoretical
knowledge of the subject.
Does this make sense?
Apache Flume for collecting the signals.
It is a distributed, reliable, and available system for efficiently collecting, aggregating and moving large amounts of log data from many different sources to a centralized data store. Easy to configure and scale. Store the data in HDFS as files using Flume.
Hive for batch queries.
Map the data files in HDFS as external tables in Hive warehouse. Write SQL like queries using HiveQL whenever you need offline-batch processing.
HBase for random real-time reads/writes.
Since HDFS, being a FS, lacks the random read/write capability, you would require a DB to serve that purpose. Looking at your use case HBase seems good to me. I would not say MongoDB or CouchDB as you are not dealing with documents here and both these are document-oriented databases.
Impala for fast, interactive queries.
Impala allows you to run fast, interactive SQL queries directly on your data stored in HDFS or HBase. Unlike Hive it does not use MapReduce. It instead leverages the power of MPP so it's good for real time stuff. And it's easy to use since it uses the same metadata, SQL syntax (Hive SQL), ODBC driver etc as Hive.
HTH
Depending on the type of analysis, CouchDB, HBase of Flume may be all be good choices. For strictly numeric "write-once" metrics data graphite is a very popular open source solution.