Hi All!

Thanks a lot for many great and prompt responses!

hi Adam, please look at some screenshots i have uploaded to Falcon-790 that
> should address #1.
>

Yes, the landing page is nice
https://issues.apache.org/jira/secure/attachment/12686141/landing.png
because it provides the search feature. It's super useful!
I wonder if this patch can be split into two smaller patches (1 for landing
page with search features and 1 for write-access to entities), if this
helps to commit a new landing page to trunk sooner.

Being able to create/edit entities in the Web UI is really great, but in
our case, it's not a priority yet. We currently use Falcon to schedule and
manage our daily/hourly ETL and KPI processes and their feeds. To do so, we
have dedicated engineers (e.g. myself), who can create these XML files and
submit them to the Falcon server with small effort. Other people just look
at Falcon Web UI.

I guess that many many people (including Hadoop admins, data engineers and
data analysts) could benefit from the features like:

> lineage (already looks nice, thanks!),
> triage (https://issues.apache.org/jira/browse/FALCON-796 as Ajay Yadav
mentioned),
> being able to quickly find any feed and process on the list
> being able to have a quickly look at Web UI to check if some KPIs failed
during a night or are still being processed....

Being able to re-run an instance of the process from the Web UI, I would
find also very useful (the CLI for that is not that friendly, because you
must provide a timestamp and so). I created a JIRA for that in case it's
useful for other people as well
https://issues.apache.org/jira/browse/FALCON-986.

 > - (5) Can you please explain it in more detail? You want only size, hdfs
   location etc. or something more?

I can find a couple of use-cases:

1. the list of recently created partitions of the feed (initially, it can
be just a link to HUE so one can easily navigate back and forth).

2. the hint for the retention period for a dataset

In this talk
http://www.slideshare.net/AdamKawa/hadoop-operations-powered-by-hadoop-hadoop-summit-2014-amsterdam
(slides 53 and 57), I talked about the "empirical retention policy" for a
dataset. This could be (more or less) calculated using the last registered
access time kept by NameNode for each file (with the precision
of dfs.namenode.accesstime.precision). Let's say, you have a dataset where
a new directory is created each day. If you aggregate the information about
the last modification time and last registered access time for the each
file in the each directory, then you can get some insight about what
retention for a dataset should be. So you know, how many days after the
creation, an instance of the dataset is not accessed anymore... I can
provide an exemplary Python code for that as I have it implemented
somewhere... :)

This hint could be helpful to set a better retention period for a feed. In
practice, many owner of the dataset don't really know what the retention
should be and they specify too long periods.

3. automatic detection for the hot datasets

In the same talk (slides 57-61), I talked about hot datasets. Initially, it
could be as simple as counting at the number of processes that depend on
that feed, and use some colours (e.g. red, yellow, green) or numbers, to
distinguish popular feeds in some way

If the dataset is really hot, you might want to temporarily increase its
replication factor for the most recent instances (to benefit from the
higher data locality). Falcon might also take care adjusting HDFS
replication factor since it's data management... :)

> *P.S. *To see how we are thinking about monitoring dashboard you can take
a
look at the POC that I did at https://github.com/ajayyadav/falcon-dashboard
Currently this works on mocked API calls and we are working to provide the
backend APIs in falcon as of now.

Thanks for sharing, I will have a look!

Cheers!
Adam

Reply via email to