Hi All,

Managed to capture the trace for a timed out query using probabilistic tracing (attached). Seems the timeout is caused by DigestMismatchException causing a global read repair. I guess this is due to dclocal_read_repair_chance = 0.0 for the table. Can someone please confirm this, and that dclocal_read_repair_chance = 0.1 will prevent this? 



On Thu, Sep 15, 2016 at 6:52 PM, Joseph Tech <jaalex.tech@gmail.com> wrote:
I added the error logs and see that the timeouts are in a range b/n 2 to 7s. Samples below: 

 Query error after 5354 ms: [4 bound values] <query>
 Query error after 6658 ms: [4 bound values] <query> 
 Query error after 4596 ms: [4 bound values] <query> 
 Query error after 2068 ms: [4 bound values] <query> 
 Query error after 2904 ms: [4 bound values] <query> 

There is no specific socket timeout set on the client side, so it would take the default of 12s. The read_request_timeout_in_ms is set to 5s. In this case, how do the errors happen in <5s ? . Is there any other factor that would cause a fail-fast scenario during the read? 



On Wed, Sep 7, 2016 at 5:26 PM, Joseph Tech <jaalex.tech@gmail.com> wrote:
Thanks, Romain for the detailed explanation. We use log4j 2 and i have added the driver logging for slow/error queries, will see if it helps to provide any pattern once in Prod. 

I tried getendpoints and getsstables for some of the timed out keys and most of them listed only 1 SSTable .There were a few which showed 2 SSTables. There is no specific trend on the keys, it's completely based on the user access, and the same keys return results instantly from cqlsh

On Tue, Sep 6, 2016 at 1:57 PM, Romain Hardouin <romainh_ml@yahoo.fr> wrote:
There is nothing special in the two sstablemetadata outuputs but if the timeouts are due to a network split or overwhelmed node or something like that you won't see anything here. That said, if you have the keys which produced the timeouts then, yes, you can look for a regular pattern (i.e. always the same keys?). 

You can find sstables for a given key with nodetool:
    nodetool getendpoints <keyspace> <cf> <key>
Then you can run the following command on one/each node of the enpoints:
    nodetool getsstables <keyspace> <cf> <key>

If many sstables are shown in the previous command it means that your data is fragmented but thanks to LCS this number should be low.

I think the most usefull actions now would be:

 1) Enable DEBUG for o.a.c.db.ConsistencyLevel, it won't spam your log file, you will see the following when errors will occur:
    - Local replicas [<endpoint1>, ...] are insufficient to satisfy LOCAL_QUORUM requirement of X live nodes in '<dc>'

    You are using C* 2.1 but you can have a look at the C* 2.2 logback.xml: https://github.com/apache/cassandra/blob/cassandra-2.2/conf/logback.xml
    I'm using it on production, it's better because it creates a separate debug.log file with a asynchronous appender.

   Watch out when enabling:
    <appender-ref ref="ASYNCDEBUGLOG" />
   Because the default logback configuration set all o.a.c in DEBUG:
        <logger name="org.apache.cassandra" level="DEBUG"/>
   Instead you can set: 
   <logger name="org.apache.cassandra" level="INFO"/>
   <logger name="org.apache.cassandra.db.ConsistencyLevel" level="DEBUG"/>

    Also, if you want to restrict debug.log to DEBUG level only (instead of DEBUG+INFO+...) you can add a LevelFilter to ASYNCDEBUGLOG in logback.xml:
        <filter class="ch.qos.logback.classic.filter.LevelFilter">

  Thus, the debug.log file will be empty unless some Consistency issues happen.
 2) Enable slow queries log at the driver level with a QueryLogger:
   Cluster cluster = ...
   // log queries longer than 1 second, see also withDynamicThreshold
   QueryLogger queryLogger = QueryLogger.builder(cluster).withConstantThreshold(1000).build();
    Then in your driver logback file:
        <logger name="com.datastax.driver.core.QueryLogger.SLOW" level="DEBUG" />
 3) And/or: you mentioned that you use DSE so you can enable slow queries logging in dse.yaml (cql_slow_log_options)



Le Lundi 5 septembre 2016 20h05, Joseph Tech <jaalex.tech@gmail.com> a écrit :

Attached are the sstablemeta outputs from 2 SSTables of size 28 MB and 52 MB (out2). The records are inserted with different TTLs based on their nature ; test records with 1 day, typeA records with 6 months, typeB records with 1 year etc. There are also explicit DELETEs from this table, though it's much lower than the rate of inserts.

I am not sure how to interpret this output, or if it's the right SSTables that were picked. Please advise. Is there a way to get the sstables corresponding to the keys that timed out, though they are accessible later.

On Mon, Sep 5, 2016 at 10:58 PM, Anshu Vajpayee <anshu.vajpayee@gmail.com> wrote:
We have seen read time out issue in cassandra due to high droppable tombstone ratio for repository. 

Please check for high droppable tombstone ratio for your repo. 

On Mon, Sep 5, 2016 at 8:11 PM, Romain Hardouin <romainh_ml@yahoo.fr> wrote:
Yes dclocal_read_repair_chance will reduce the cross-DC traffic and latency, so you can swap the values ( https://issues.apache.org/ji ra/browse/CASSANDRA-7320 ). I guess the sstable_size_in_mb was set to 50 because back in the day (C* 1.0) the default size was way too small: 5 MB. So maybe someone in your company tried "10 * the default" i.e. 50 MB. Now the default is 160 MB. I don't say to change the value but just keep in mind that you're using a small value here, it could help you someday.

Regarding the cells, the histograms shows an *estimation* of the min, p50, ..., p99, max of cells based on SSTables metadata. On your screenshot, the Max is 4768. So you have a partition key with ~ 4768 cells. The p99 is 1109, so 99% of your partition keys have less than (or equal to) 1109 cells. 
You can see these data of a given sstable with the tool sstablemetadata.



Le Lundi 5 septembre 2016 15h17, Joseph Tech <jaalex.tech@gmail.com> a écrit :

Thanks, Romain . We will try to enable the DEBUG logging (assuming it won't clog the logs much) . Regarding the table configs, read_repair_chance must be carried over from older versions - mostly defaults. I think sstable_size_in_mb was set to limit the max SSTable size, though i am not sure on the reason for the 50 MB value.

Does setting dclocal_read_repair_chance help in reducing cross-DC traffic (haven't looked into this parameter, just going by the name).

By the cell count definition : is it incremented based on the number of writes for a given name(key?) and value. This table is heavy on reads and writes. If so, the value should be much higher?

On Mon, Sep 5, 2016 at 7:35 AM, Romain Hardouin <romainh_ml@yahoo.fr> wrote:

Try to put org.apache.cassandra.db. ConsistencyLevel at DEBUG level, it could help to find a regular pattern. By the way, I see that you have set a global read repair chance:
    read_repair_chance = 0.1
And not the local read repair:
    dclocal_read_repair_chance = 0.0
Is there any reason to do that or is it just the old (pre 2.0.9) default configuration? 

The cell count is the number of triplets: (name, value, timestamp)

Also, I see that you have set sstable_size_in_mb at 50 MB. What is the rational behind this? (Yes I'm curious :-) ). Anyway your "SSTables per read" are good.



Le Lundi 5 septembre 2016 13h32, Joseph Tech <jaalex.tech@gmail.com> a écrit :

Hi Ryan,

Attached are the cfhistograms run within few mins of each other. On the surface, don't see anything which indicates too much skewing (assuming skewing ==keys spread across many SSTables) . Please confirm. Related to this, what does the "cell count" metric indicate ; didn't find a clear explanation in the documents.


On Thu, Sep 1, 2016 at 6:30 PM, Ryan Svihla <rs@foundev.pro> wrote:
Have you looked at cfhistograms/tablehistograms your data maybe just skewed (most likely explanation is probably the correct one here)


Ryan Svihla

From: Joseph Tech <jaalex.tech@gmail.com>
Sent: Wednesday, August 31, 2016 11:16 PM
Subject: Re: Read timeouts on primary key queries
To: <user@cassandra.apache.org>


The desc table is below (only col names changed) : 

    id1 text,
    id2 text,
    id3 text,
    id4 text,
    f1 text,
    f2 map<text, text>,
    f3 map<text, text>,
    created timestamp,
    updated timestamp,
    PRIMARY KEY (id1, id2, id3, id4)
    AND bloom_filter_fp_chance = 0.01
    AND caching = '{"keys":"ALL", "rows_per_partition":"NONE"}'
    AND comment = ''
    AND compaction = {'sstable_size_in_mb': '50', 'class': 'org.apache.cassandra.db. compaction. LeveledCompactionStrategy'}
    AND compression = {'sstable_compression': 'org.apache.cassandra.io. compress.LZ4Compressor'}
    AND dclocal_read_repair_chance = 0.0
    AND default_time_to_live = 0
    AND gc_grace_seconds = 864000
    AND max_index_interval = 2048
    AND memtable_flush_period_in_ms = 0
    AND min_index_interval = 128
    AND read_repair_chance = 0.1
    AND speculative_retry = '99.0PERCENTILE';

and the query is select * from tbl where id1=? and id2=? and id3=? and id4=?

The timeouts happen within ~2s to ~5s, while the successful calls have avg of 8ms and p99 of 15s. These times are seen from app side, the actual query times would be slightly lower. 

Is there a way to capture traces only when queries take longer than a specified duration? . We can't enable tracing in production given the volume of traffic. We see that the same query which timed out works fine later, so not sure if the trace of a successful run would help.


On Wed, Aug 31, 2016 at 8:05 PM, Patrick McFadin <pmcfadin@gmail.com> wrote:
If you are getting a timeout on one table, then a mismatch of RF and node count doesn't seem as likely. 

Time to look at your query. You said it was a 'select * from table where key=?' type query. I would next use the trace facility in cqlsh to investigate further. That's a good way to find hard to find issues. You should be looking for clear ledge where you go from single digit ms to 4 or 5 digit ms times. 

The other place to look is your data model for that table if you want to post the output from a desc table.


On Tue, Aug 30, 2016 at 11:07 AM, Joseph Tech <jaalex.tech@gmail.com> wrote:
On further analysis, this issue happens only on 1 table in the KS which has the max reads. 

@Atul, I will look at system health, but didnt see anything standing out from GC logs. (using JDK 1.8_92 with G1GC). 

@Patrick , could you please elaborate the "mismatch on node count + RF" part.

On Tue, Aug 30, 2016 at 5:35 PM, Atul Saroha <atul.saroha@snapdeal.com> wrote:
There could be many reasons for this if it is intermittent. CPU usage + I/O wait status. As read are I/O intensive, your IOPS requirement should be met that time load. Heap issue if CPU is busy for GC only. Network health could be the reason. So better to look system health during that time when it comes.

------------------------------ ------------------------------ ------------------------------ ---------------------------
Atul Saroha

Lead Software Engineer
M: +91 8447784271 T: +91 124-415-6069 EXT: 12369
Plot # 362, ASF Centre - Tower A, Udyog Vihar,
 Phase -4, Sector 18, Gurgaon, Haryana 122016, INDIA

On Tue, Aug 30, 2016 at 5:10 PM, Joseph Tech <jaalex.tech@gmail.com> wrote:
Hi Patrick,

The nodetool status shows all nodes up and normal now. From OpsCenter "Event Log" , there are some nodes reported as being down/up etc. during the timeframe of timeout, but these are Search workload nodes from the remote (non-local) DC. The RF is 3 and there are 9 nodes per DC.


On Mon, Aug 29, 2016 at 11:07 PM, Patrick McFadin <pmcfadin@gmail.com> wrote:
You aren't achieving quorum on your reads as the error is explains. That means you either have some nodes down or your topology is not matching up. The fact you are using LOCAL_QUORUM might point to a datacenter mis-match on node count + RF. 

What does your nodetool status look like?


On Mon, Aug 29, 2016 at 10:14 AM, Joseph Tech <jaalex.tech@gmail.com> wrote:

We recently started getting intermittent timeouts on primary key queries (select * from table where key=<key>)

The error is : com.datastax.driver.core.excep tions.ReadTimeoutException: Cassandra timeout during read query at consistency LOCAL_QUORUM (2 responses were required but only 1 replica
a responded)

The same query would work fine when tried directly from cqlsh. There are no indications in system.log for the table in question, though there were compactions in progress for tables in another keyspace which is more frequently accessed. 

My understanding is that the chances of primary key queries timing out is very minimal. Please share the possible reasons / ways to debug this issue. 

We are using Cassandra 2.1 (DSE 4.8.7).