1. Yes, in my mind I was imagining a formal grammar specified by antlr, like we use in stellar with all the accompanying tooling that comes from an antlr grammar.
2 and 3 are worth pondering. I'll reserve comment until I've pondered. ;) On Tue, Jan 31, 2017 at 2:19 PM, Matt Foley <[email protected]> wrote: > Casey, this sounds great. > > 1. I think you have a model for how to do the natural-language-look DSL, > that maybe isn’t clear to the rest of us. Does your contemplated approach > meet the following? > a) Can it be specified as a formal grammar, not too complex, so anyone who > bothers can visually parse a sentence and see that it is conforming (or > not)? – Obviously I’m concerned that we not bite off a full NLP problem for > the sake of ease of use. > b) Can it dump a human-understandable parse tree on request, so if it does > the unexpected (or just fails to work), the user can easily figure out what > the program did and why? > c) And you already said it wouldn’t be much more complex to implement than > using a Map parameter list. > > Given those 3 things, I’d vote for the NL, otherwise the parameter list. > > 2. One more thing I’d recommend for PROFILE_LOOKBACK(): wildcarding of > Groups. > - Easy to implement for groups based on small finite sets like day-of-week. > - Also easy altho perhaps slow for large finite sets, and > multi-dimensional group sub-keys. > - Not clear how to do for indeterminate sets from random Stellar > functions, but we should talk to HBase gurus and see what we can do, using > HBase’s scan capability. > - If groups sub-keys are not already at the end of the row key structure, > we may want to move them there, to make the latter case easier and more > efficient. > > 3. There’s another thing I’ve been noodling, that deserves its own > discussion, but is relevant to mention here – or if you really like it, > could be included in PROFILE_LOOKBACK project, as it’s not a difficult > thing: > Currently the Profiler config settings, at the time a Profile is run, get > burned into the HBase row keys, such that if you don’t _a priori_ know > those config settings, you can’t read the Profile. I’d like to suggest we > start saving Profile metadata, also in HBase, every time a new Profile is > started (and, if possible, ended) so that I can say “Get that old profile > from November 14-21, whatever its metadata was” without needing to know its > period, groups, etc. Obviously it would be nice to be able to query the > metadata itself, too, but just having the metadata gives us the right start. > > --Matt > > On 1/31/17, 10:43 AM, "Casey Stella" <[email protected]> wrote: > > Regarding the "?" syntax: > Wouldn't that be forking cron syntax so now we have a metron cron? If > we're constructing our own syntax, then why not do it so that it reads > like > natural language? > > Regarding the holiday problem: > Agreed, it's a smaller problem than constructing a DSL, but that's not > really the point, I think. The concern is that it would be unable to be > expressed using cron syntax in a natural way without modifying cron > syntax, > which would be constructing a new DSL. If quartz has a clever way of > doing > that, then I'd like to see it. From a quick search, I haven't seen a > scheduling example with a compact syntax that shows skipping holidays > with > cron syntax. > > > On Tue, Jan 31, 2017 at 1:29 PM, Nick Allen <[email protected]> > wrote: > > > > > > > - Cron syntax allows you to construct only absolute lookbacks > (i.e. > > > "every tuesday at 3PM" not "every tuesday at the current hour") > > > > > > I think Cron would work for this. I am no expert on cron > expressions, but > > I think the following examples would work. > > > > - If you want "every Tuesday at 3 PM" > > - 0 0 15 ? * TUE * > > - If you want "every Tuesday at current hour" then use something > like > > the "?" placeholder maybe. > > - 0 0 ? ? * TUE * > > > > - Cron syntax allows you to specify a point in time, not a > duration. We > > > could, of course, specify a duration as another argument > > > > > > Yes, a separate argument would be necessary. We would have to allow > the > > user to specify either a "start from date/time" or the "number of > intervals > > to look back". > > > > Cron syntax does not allow you to skip things like holidays, etc. > > > > > > I agree, out-of-the-box Cron does not solve holiday calendars. But > this > > would be a smaller problem to solve then creating our own DSL. > > > > There is a tradition of creating shortcuts that look something like > @Daily > > or @Weekdays or @Tuesdays that we could also use to make things > easier for > > users. > > > > I have used Quartz with cron expressions in the past and there was > some way > > to handle holidays with that. I think you could create a custom > calendar > > for the holidays and call it something; aka @USHolidays. And then > you > > would say "every Tuesday" except @USHolidays or something like > that. I'd > > have to look into this some more. > > > > And there are also nice online Cron expression "translators" that we > could > > mimic in a Metron user interface. For example, https://crontab.guru > . > > > > > > > > > > On Tue, Jan 31, 2017 at 12:00 PM, Casey Stella <[email protected]> > wrote: > > > > > I actually did consider cron initially but dismissed it for the > following > > > reasons: > > > > > > - Cron syntax allows you to construct only absolute lookbacks > (i.e. > > > "every tuesday at 3PM" not "every tuesday at the current hour") > > > - Cron syntax allows you to specify a point in time, not a > duration. > > We > > > could, of course, specify a duration as another argument > > > - Cron syntax does not allow you to skip things like holidays, > etc. > > > > > > You could use Cron syntax as part of a broader API to specify the > days to > > > look back and have other arguments handle the aspects that cron > doesn't > > > support out of the box. I share your concern at making another > DSL, but > > > cron seemed to not be a complete solution and it's syntax, despite > being > > > well known by admins, may not be well known to analysts. Also, > and this > > is > > > just a personal bias, I find it inscrutable without a fair amount > of > > > wikipedia and man page reading. > > > > > > On Tue, Jan 31, 2017 at 11:47 AM, Nick Allen <[email protected]> > wrote: > > > > > > > I do prefer the flexibility of the DSL, but would prefer not to > create > > > yet > > > > another DSL for our users to learn. Couldn't we somehow use cron > > > > expressions for this functionality? > > > > > > > > On Mon, Jan 23, 2017 at 3:01 PM, Casey Stella < > [email protected]> > > > wrote: > > > > > > > > > Hi All, > > > > > > > > > > I'm planning to expand the capabilities of PROFILE_GET and > wanted to > > > pass > > > > > an idea past the community. > > > > > > > > > > *Current State* > > > > > > > > > > Currently, the functionality of PROFILE_GET is fairly > > straightforward: > > > > > > > > > > - profile - The name of the profile. > > > > > - entity - The name of the entity. > > > > > - durationAgo - How long ago should values be retrieved > from? > > > > > - units - The units of 'durationAgo'. > > > > > - groups_list - Optional, must correspond to the 'groupBy' > list > > used > > > > in > > > > > profile creation - List (in square brackets) of groupBy > values > > used > > > to > > > > > filter the profile. Default is the empty list, meaning > groupBy was > > > not > > > > > used > > > > > when creating the profile. > > > > > - config_overrides - Optional - Map (in curly braces) of > > name:value > > > > > pairs, each overriding the global config parameter of the > same > > name. > > > > > Default is the empty Map, meaning no overrides. > > > > > > > > > > This has the advantage of providing a relatively simple > mechanism to > > > > > support the dominant use-case, gathering the profiles for a > trailing > > > > > window. The issues, however, are a couple: > > > > > > > > > > - We may need more complex semantics for specifying the > window > > > > > (motivated below) > > > > > - As such, this couples the gathering of the profiles with > the > > > > > specification of the window. > > > > > > > > > > I propose to decouple these two concepts. I propose that we > extract > > the > > > > > notion of the lookback into a separate, more featureful > function > > called > > > > > PROFILE_LOOKBACK() which could be composed with an adjusted > > > PROFILE_GET, > > > > > whose arguments look like: > > > > > > > > > > > > > > > - profile - The name of the profile. > > > > > - entity - The name of the entity. > > > > > - timestamps - The list of timestamps to retrieve > > > > > - groups_list - Optional, must correspond to the 'groupBy' > list > > used > > > > in > > > > > profile creation - List (in square brackets) of groupBy > values > > used > > > to > > > > > filter the profile. Default is the empty list, meaning > groupBy was > > > not > > > > > used > > > > > when creating the profile. > > > > > - config_overrides - Optional - Map (in curly braces) of > > name:value > > > > > pairs, each overriding the global config parameter of the > same > > name. > > > > > Default is the empty Map, meaning no overrides. > > > > > > > > > > So, PROFILE_GET would have the output of PROFILE_LOOKBACK > passed to > > it > > > as > > > > > its 3rd argument (e.g. PROFILE_GET( 'my_profile', 'my_entity', > > > > > PROFILE_LOOKBACK(...)) ). > > > > > > > > > > *Motivation for Change* > > > > > > > > > > The justification for this is that sometimes you want to > compare time > > > > bins > > > > > for a long duration back, but you don't want to skew the data > by > > > > including > > > > > periods that aren't distributionally similar (due to seasonal > data, > > for > > > > > instance). You might want to compare a value to statistically > > baseline > > > > of > > > > > the median of the values for the same time window on the same > day for > > > the > > > > > last month (e.g. every tuesday at this time). > > > > > > > > > > Also, we might want a trailing window that does not start at > the > > > current > > > > > time (in wall-clock), but rather starts an hour back or from > the time > > > > that > > > > > the data was originally ingested. > > > > > > > > > > > > > > > *PROFILE_LOOKBACK* > > > > > > > > > > I propose that we support the following features: > > > > > > > > > > - A starting point that is not current time > > > > > - Sparse bins (i.e. the last hour for every tuesday for the > last > > > > month) > > > > > - The ability to skip events (e.g. weekends, holidays) > > > > > > > > > > > > > > > This would result in a new function with the following > arguments: > > > > > > > > > > - > > > > > > > > > > from - The lookback starting point (default to now) > > > > > - > > > > > > > > > > fromUnits - The units for the lookback starting point > > > > > - > > > > > > > > > > to - The ending point for the lookback window (default to > from + > > > > > binSize) > > > > > - > > > > > > > > > > toUnits - The units for the lookback ending point > > > > > - > > > > > > > > > > including - A list of conditions which we would skip. > > > > > - weekend > > > > > - holiday > > > > > - sunday through saturday > > > > > - > > > > > > > > > > excluding - A list of conditions which we would skip. > > > > > - weekend > > > > > - holiday > > > > > - sunday through saturday > > > > > - > > > > > > > > > > binSize - The size of the lookback bin > > > > > - > > > > > > > > > > binUnits - The units of the lookback bin > > > > > > > > > > Given the number of arguments and their complexity and the > fact that > > > > many, > > > > > many are optional, I propose that either > > > > > > > > > > - PROFILE_LOOKBACK take a Map so that we can get > essentially named > > > > > params in stellar. > > > > > - PROFILE_LOOKBACK accept a string backed by a DSL to > express > > these > > > > > criteria > > > > > > > > > > > > > > > Ok, so that's a lot to take in. How about we look at some > motivating > > > > > use-cases. > > > > > > > > > > *Base Case: A lookback of 1 hour ago* > > > > > > > > > > As a map, this would look like: > > > > > > > > > > PROFILE_LOOKBACK( { 'binSize' : 1, 'binUnits' : 'HOURS' } ) > > > > > > > > > > As a DSL this would look like: > > > > > PROFILE_LOOKBACK( '1 hour bins from now') > > > > > > > > > > > > > > > *The same time window every tuesday for the last month > starting one > > > hour > > > > > ago* > > > > > > > > > > Just to make this as clear as possible, if this is run at 3PM > on > > Monday > > > > > January 23rd, 2017, it would include the following bins: > > > > > > > > > > - January 17th, 2PM - 3PM > > > > > - January 10th, 2PM - 3PM > > > > > - January 3rd, 2PM - 3PM > > > > > - December 27th, 2PM - 3PM > > > > > > > > > > As a map, this would look like: > > > > > > > > > > PROFILE_LOOKBACK( { 'from' : 1, 'fromUnits' : 'HOURS', 'to' : > 1, > > > > 'toUnits' > > > > > : 'MONTH', 'including' : [ 'tuesday' ], 'binSize' : 1, > 'binUnits' : > > > > 'HOURS' > > > > > } ) > > > > > > > > > > As a DSL this would look like: > > > > > PROFILE_LOOKBACK( '1 hour bins from 1 hour to 1 month including > > > > tuesdays') > > > > > > > > > > *The same time window every sunday for the last month starting > one > > hour > > > > ago > > > > > skipping holidays* > > > > > > > > > > Just to make this as clear as possible, if this is run at 3PM > on > > Monday > > > > > January 22rd, 2017, it would include the following bins: > > > > > > > > > > - January 16th, 2PM - 3PM > > > > > - January 9th, 2PM - 3PM > > > > > - January 2rd, 2PM - 3PM > > > > > - NOT December 25th > > > > > > > > > > As a map, this would look like: > > > > > > > > > > PROFILE_LOOKBACK( { 'from' : 1, 'fromUnits' : 'HOURS', 'to' : > 1, > > > > 'toUnits' > > > > > : 'MONTH', 'including' : [ 'tuesday'], 'excluding' : [ > 'holidays' ], > > > > > 'binSize' : 1, 'binUnits' : 'HOURS' } ) > > > > > > > > > > As a DSL this would look like: > > > > > PROFILE_LOOKBACK( '1 hour bins from 1 hour to 1 month including > > > tuesdays > > > > > excluding holidays') > > > > > > > > > > *DSL vs API* > > > > > > > > > > So, here's my personal rundown of the two approaches: > > > > > > > > > > DSL: > > > > > > > > > > - PRO > > > > > - Clear. As you can see, it reads like a sentence > > > > > - Concise > > > > > - CON: > > > > > - More complex to implement > > > > > - Another DSL to learn > > > > > > > > > > API: > > > > > > > > > > - PRO > > > > > - Simpler to implement (though marginally so, IMO) > > > > > - CON > > > > > - A bit more complex to understand (also, IMO) > > > > > > > > > > I'd like to solicit feedback from the community at this point: > > > > > > > > > > - What do you think of this change? > > > > > - Would you prefer the DSL, API or other approach? > > > > > > > > > > Thanks, > > > > > > > > > > Casey > > > > > > > > > > > > > > > > > > > > > -- > > > > Nick Allen <[email protected]> > > > > > > > > > > > > > > > -- > > Nick Allen <[email protected]> > > > > > > >
