SDK Starfish v0 requirements

[AbhiPrasad]

Add cache hit/miss rate

Any call to cache that doesn't return data (null) is treated as a miss. We define a span that has a cache hit as having the span data field cache.hit as true. Might require SDK to patch cache abstractions of framework (django).

To be done in Python SDK for any frameworks that support this - django is required.

Cache hit/miss rate

Preview Give feedback

1. feat: Add span data for db.system and cache.hit develop#917
   +0
   
2. Add cache.hit data to Django spans sentry-python#2043
   Enhancement Integration: Django Status: Backlog +3
   

Add db platform data to to span data

Add this information to span data field as db.system, matching OpenTelemetry's well known conventions.

For example, db.system of postgresql would indicate that it's a postgres database.

DB platform data.

Preview Give feedback

1. feat: Add span data for db.system and cache.hit develop#917
   +0
   
2. feat: Add db.system indicators to redis, sqlalchemy, and django sentry-python#2035
   +0
   
3. feat(tracing): Add db.system span data to DB spans sentry-javascript#7952
   +0
   
4. Add db.system constant sentry-python#2037
   +0
   
5. feat(redis): Add db.system to redis db spans sentry-python#2038
   +0
   
6. feat: Add db.system to sqlalchemy spans sentry-python#2039
   +0
   
7. feat(django): Add db.system sentry-python#2040
   +0
   
8. Add db.system to the span data for mongo db. sentry-python#2042
   +0
   

Add cache item size

Add a field to cache spans that defines how big the item that is being get/set: cache.item_size. This is an integer and should be in bytes. Blocked by cache hit/miss rate work.

Cache Item size

Preview Give feedback

1. Added cache.item_size to span data conventions develop#933
   +0
2. Add cache.item_size to Python SDK

Add information used for action field

The action field on the extracted span metrics is either the operation for DB spans (SELECT/DELETE etc.), and http method for HTTP spans (GET/PUT etc.).

For operation, we'll be using and setting db.operation on span data. For http method, we'll be using http.method on span data.

If db.operation OR http.method is not set on span data, Relay will have to parse it out from the span description.

Action

Preview Give feedback

1. feat: Add conventions for http.method and db.operation develop#924
   +0
   
2. ref: Use constants for http.query and http.fragment sentry-python#2053
   +0
   
3. feat: Use http.method instead of method sentry-python#2054
   +0
   
4. feat(tracing): Use http.method for span data sentry-javascript#7990
   +0
   
5. feat(node): Add http.method to node http spans sentry-javascript#7991
   +0
   

[wmak]

From the Slack thread we'll also need:

• Action
  • This would be the operation for DB spans (SELECT/DELETE etc.), and http method for HTTP spans (GET/PUT etc.)
  • There was discussion where we're not sure if this should be done in the SDK or in ingestion
• Span Status
  • The main motivation of this is so we can get client timeouts for http spans

[AbhiPrasad]

We can add all of the db call-level attributes if possible: https://opentelemetry.io/docs/reference/specification/trace/semantic_conventions/database/#call-level-attributes: db.operation.

db.operation in this case would be the field that would eventually be extracted to be action. Note that we can only provide db.operation if the underlying framework gives us this information, otherwise it will have to be extracted from the query - which has to be done server-side.

I'm still confused about span status - we attach http status codes under span.status already on the SDK. Can we not check for a 504 or a 408 and label timeout appropriately? Why does this require SDK changes?

[mitsuhiko]

What is a client timeout here? Presumably if something times out, you never get a response and thus no status code. What do SDKs report in that case? I wouldn't be surprised if a timeout usually translates into some sort of socket error which requires custom handling and might either result in a sentry error (and failed transaction), or is manually handled and the output is undefined.

[AbhiPrasad]

Based on a slack thread - here's a snippet from @gggritso about behaviour that he would like to see for timeouts.

import urllib
import urllib.request

try:
    response = urllib.request.urlopen('http://python.org/', timeout=0.0001)
    html = response.read()
except urllib.error.URLError as e:
    print("URL Error", e.reason)
    if (e.reason == 'timed out'):
        span['status'] = 'time out'
        raise e

[AbhiPrasad]

Given we've released the Python SDK with cache hit information, I think we can say we have enough data for v0 of starfish.

action (db operation, http method) is going to be parsed by Relay for the most part (but both Node/Python should be sending them regardless)

span status is still an open concern, but I think we can defer that will v1. cc @alexjillard

[smeubank]

the SDK should be sending http.method where possible under span data. In both node/python it was determined that none of the underlying instrumentation could parse and determine db.operation so this was something relay needed to do.

So Abhi doesn't have to sign into GH while OOO