Hi,

Apologies for a dev related hijack. We can continue the dev related discussion in JIRA or the dev list.

Seeing this discussion makes me think that Benedict and Ryan you will both have opinions about https://issues.apache.org/jira/browse/CASSANDRA-12372 (remove memtable_cleanup_threshold, not done yet) and the already committed https://issues.apache.org/jira/browse/CASSANDRA-12228 (deprecate MCT and change yaml documentation) which are both directly related to the defaults and advice we give in cassandra.yaml for MCT.

Ariel
On Thu, Sep 8, 2016, at 08:38 AM, Benedict Elliott Smith wrote:
The thing is, this isn't about opinions.  I don't really want to get into an argument either, but characterising my statements as opinion does invite me to respond with their factual basis...  The only legitimate opinions here would be around the prevalence of cluster characteristics, of which we both have anecdata (and of which your anecdata is no doubt a larger sample than mine, but probably no less biased a sample).

To start, I will note I have already indicated that I believe a majority of clusters are likely to become compaction-constrained, and as a result the defaults may not be optimal.  However:

1) There are definitely workloads that are not compaction constrained, and these clusters definitely do not want the values you suggest.  Cassandra is best used for bursty workloads, where a cluster is briefly saturated and has periods of relative calm.  In these workloads, there is plenty of time to make up compaction later, and is really the whole benefit of LSMTs that C* offers.  In these systems, a lower value is definitely preferable.

2) Similarly, in clusters with DTCS/TWCS, heavy-TTL use, or simply very small datasets, compaction costs are significantly curtailed, and a lower value can easily be preferred to increase cluster throughput.

3) Finally, if you have one data table that receives a majority of your writes, a value of 0.6 will be liable to cause write stalls because only one memtable can be flushing at any moment.  Once you have more than one data table receiving a significant proportion of writes this becomes a non-issue, but since Cassandra is optimal with only one main data table, and fewer tables are encouraged, this is a really important point for a subset of users.  These users should definitely not use your suggested value, and should cap at 0.49 or less.

I will reiterate for the umpteenth time I'm not contradicting your statements that the defaults can be improved or that your suggestion is better for (probably) the majority of users.  But it is not a universal truth, and the blanket statement may encourage a proliferation of bad configs.

But I will also reiterate the documented advice in the yaml is indeed atrocious, and offers much more misleading advice than yours, so we should perhaps try to distill this discussion into a more palatable snippet for the yaml.






On 29 August 2016 at 16:22, Ryan Svihla <rs@foundev.pro> wrote:
Benedict I really don't want to turn this into a battle about who's opinion is more valid and I really respect all the good work you've done for Apache Cassandra.

I'll just reiterate that I'm comfortable saying 0.6 is a good starting point and it is often not the ideal once you go through more thorough testing, all of which I said initially and I still think is a reasonable statement.

-regards,

Ryan Svihla





On Sat, Aug 27, 2016 at 9:31 AM -0500, "Benedict Elliott Smith" <benedict@apache.org> wrote:


I did not claim you had no evidence, only that your statement lacked justification.  Again, nuance is important.

I was suggesting that blanket statements without the necessary caveats, to the user mailing list, countermanding the defaults without 'justification' (explanation, reasoning) is liable to cause confusion on what best practice is.  I attempted to provide some of the missing context to minimise this confusion while still largely agreeing with you.

However you should also bear in mind that you work as a field engineer for DataStax, and as such your sample of installation behaviours will be biased - towards those where the defaults have not worked well.



On Saturday, 27 August 2016, Ryan Svihla <rs@foundev.pro> wrote:
 I have been trying to get the docs fixed for this for the past 3 months, and there already is a ticket open for changing the defaults. I don't feel like I've had a small amount of evidence here. All observation in the 3 years of work in the field suggests compaction keeps coming up as the bottleneck when you push Cassandra ingest.
0.6 as an initial setting has fixed 20+ broken clusters in practice and it improved overall performance in every case from defaults of 0.33 to defaults of 0.03 (yaml suggests per core flush writers, add in the prevelance of HT and you see a lot of 24+ flush writer systems in the wild)

No disrespect intended but that default hasn't worked out well at all in my exposure to it, and 0.6 has never been worse than the default yet. Obviously write patterns, heap configuration, memtable size limits and what not affect the exact optimal setting and I've rarely had it end up 0.6 after a tuning exercise. I never intended that as a blanket recommendation, just a starting one.

_____________________________
From: Benedict Elliott Smith <benedict@apache.org>
Sent: Friday, August 26, 2016 9:40 AM
Subject: Re: Guidelines for configuring Thresholds for Cassandra metrics


The default when I wrote it was 0.4 but it was found this did not saturate flush writers in JBOD configurations. Iirc it now defaults to 1/(1+#disks) which is not a terrible default, but obviously comes out much lower if you have many disks.

This smaller value behaves better for peak performance, but in a live system where compaction is king not saturating flush in return for lower write amplification (from flushing larger memtables) will indeed often be a win.

0.6, however, is probably not the best default unless you have a lot of tables being actively written to, in which case even 0.8 would be fine. With a single main table receiving your writes at a given time, 0.4 is probably an optimal value, when making this trade off against peak performance.

Anyway, it's probably better to file a ticket to discuss defaults and documentation than making a statement like this without justification. I can see where you're coming from, but it's confusing for users to have such blanket guidance that counters the defaults.  If the defaults can be improved (which I agree they can) it's probably better to do that, along with better documentation, so the nuance is accounted for.


On Friday, 26 August 2016, Ryan Svihla <rs@foundev.pro> wrote:

Forgot the most important thing. Logs
ERROR you should investigate
WARN you should have a list of known ones. Use case dependent. Ideally you change configuration accordingly.
*PoolCleaner (slab or native) - good indication node is tuned badly if you see a ton of this. Set memtable_cleanup_threshold to 0.6 as an initial attempt to configure this correctly.  This is a complex topic to dive into, so that may not be the best number, it'll likely be better than the default, why its not the default is a big conversation.
There are a bunch of other logs I look for that are escaping me at present but that's a good start

-regards,

Ryan Svihla




On Fri, Aug 26, 2016 at 7:21 AM -0500, "Ryan Svihla" <rs@foundev.pro> wrote:

Thomas,

Not all metrics are KPIs and are only useful when researching a specific issue or after a use case specific threshold has been set.

The main "canaries" I monitor are:
* Pending compactions (dependent on the compaction strategy chosen but 1000 is a sign of severe issues in all cases)
* dropped mutations (more than one I treat as a event to investigate, I believe in allowing operational overhead and any evidence of load shedding suggests I may not have as much as I thought)
* blocked anything (flush writers, etc..more than one I investigate)
* system hints ( More than 1k I investigate)
* heap usage and gc time vary a lot by use case and collector chosen, I aim for below 65% usage as an average with g1, but this again varies by use case a great deal. Sometimes I just looks the chart and query patterns and if they don't line up I have to do other deeper investigations
* read and write latencies exceeding SLA is also use case dependent. Those that have none I tend to push towards p99 with a middle end SSD based system having 100ms and a spindle based system having 600ms with CL one and assuming a "typical" query pattern (again query patterns and CL so vary here)
* cell count and partition size vary greatly by hardware and gc tuning but I like to in the absence of all other relevant information like to keep cell count for a partition below 100k and size below 100mb. I however have many successful use cases running more and I've had some fail well before that. Hardware and tuning tradeoff a shift this around a lot.
There is unfortunately as you'll note a lot of nuance and the load out really changes what looks right (down to the model of SSDs I have different expectations for p99s if it's a model I haven't used before I'll do some comparative testing).

The reason so much of this is general and vague is my selection bias. I'm brought in when people are complaining about performance or some grand systemic crash because they were monitoring nothing. I have little ability to change hardware initially so I have to be willing to allow the hardware to do the best it can an establish levels where it can no longer keep up with the customers goals. This may mean for some use cases 10 pending compactions is an actionable event for them, for another customer 100 is. The better approach is to establish a baseline for when these metrics start to indicate a serious issue is occurring in that particular app. Basically when people notice a problem, what did these numbers look like in the minutes, hours and days prior? That's the way to establish the levels consistently.

Regards,

Ryan Svihla







On Fri, Aug 26, 2016 at 4:48 AM -0500, "Thomas Julian" <thomasjulian@zoho.com> wrote:

Hello,

I am working on setting up a monitoring tool to monitor Cassandra Instances. Are there any wikis which specifies optimum value for each Cassandra KPIs?
For instance, I am not sure,
  1. What value of "Memtable Columns Count" can be considered as "Normal". 
  2. What value of the same has to be considered as "Critical".
I knew threshold numbers for few params, for instance any thing more than zero for timeouts, pending tasks should be considered as unusual. Also, I am aware that most of the statistics' threshold numbers vary in accordance with Hardware Specification, Cassandra Environment Setup. But, what I request here is a general guideline for configuring thresholds for all the metrics.

If this has been already covered, please point me to that resource. If anyone on their own interest collected these things, please share.

Any help is appreciated.

Best Regards,
Julian.