Capturing from chat:

what we really want from this is to not need to worry any more about what metric we put in and still get correct durations.

• Let's worry about memory consumption later - let's see how useful this feature turns out first before we optimize.
• The duration data needs to support turn penalties. If we do time-as-weight routing, we need to support them there too. We can store them at the end of the GetUncompressedWeights array, it'll add 4 bytes for each compressed geometry.
• Let's only support a weight_per_distance value from the user, rather than arbitrary total values. Everything gets scaled by the length of the road.
• The way_function can return two values per way - speed and weight_per_distance (note that both these fields are per-distance measurements).
• The turn_function can return two values - time_penalty and weight_penalty

For traffic data imports, let's look at some use cases:

• if we're doing fastest-time routing, we will need to update both the routing weight and the stored durations
• if we're doing shortest-distance routing, but adjusting durations for traffic, we need to update the stored durations, but not the weight
• if we're trying to support "minimize toll costs" in a situation where tolls are changing constantly (dynamic pricing), but without traffic data, we would need to update the weights, but not the durations

The last case is a bit of a stretch, I couldn't think of a better example. But the point is that we probably need to be able to update both with the CSV import. I think we could do:

• update the CSV file to have 4 columns, nodeA,nodeB,speed,weight_per_distance
• supply a second CSV file with nodeA,nodeB,weight_per_distance

I think if we simply don't supply data, we can efficiently skip either the duration and/or weight updates.

wow, this has been a long time coming! it will be very useful for bike routing

nodeA,nodeB,speed,weight_per_distance

We can update the data with regard to the names given in the csv header column.

Things to do here:

• Split interleaved CompressedGeometryContainer serialization into separate vectors
• Only load the duration vector if weight doesn't contain the same values (safes memory consumption)
• Make duration values in osrm-contract updatable

Hi, i have almost finished rebase of the PR to master, but some parts are missing. If it is interesting, i can create a duplicate PR with rebased patches.

@oxidase Let me just give your contributor access so you can push directly to this branch.

Please review a forced rebase of the branch to master. I have modified some commits, changed some properties in profiles and changed TurnPenalty to int16_t to make all tests green.

@TheMarex updated and rebased the brach, but in some places i did not make changes:

• traffic_light_penalty is still used in turn_function to have old behavior, because TODO fix is required
• MAX_WEIGHT_NAME_LENGTH is still a value in enum
• forward_heap_3 and reverse_heap_3 are still in use
• FIXME is still there but now related only the situation when source and target phantoms on one edge

traffic_light_penalty is still used in turn_function to have old behavior, because TODO fix is required

You are right, I confused myself with my own changelog entry. Currently it is:

• traffic_light_penalty is deprecated. traffic light penalties now need to be set over in the node function in result.weight_penalty and result.duration_penalty fields

Correct however would be:

• properties.traffic_light_penalty is deprecated. Traffic light penalties now need to be set over in the turn function. Each turn with a traffic light is marked with ExtractionTurn::has_traffic_light = true.

This looks good to me. I think we can merge this (wow so weird).

looks like std::size_t is not integral type? will change to static constexpr int MAX_WEIGHT_NAME_LENGTH = 255;

Thanks for all the effort you are putting into this.
This pull request getting merged is really great for us.

@djschilling @emiltin and anyone else who has been watching this ticket - we've merge into master but feedback on this feature would be really helpful. Anyone who has the time to test it out, please do and open tickets with any problems/confusion you encounter.

We are using an earlier state of the pull request since the start of the year and it does exactly what it should.
We use it to correctly handle access:destination tags, by putting a high weight to access:destination ways.

One thing that we encountered is a performance loose from about 40%. Its not critical for us, but i think its worth mentioning.
I have no time now to test the currently merged version, so feedback from someone else would probably be good.