Getting 408 Request Timeout when calling TimeGen-1 timegpt

Question

Getting 408 Request Timeout when calling TimeGen-1 timegpt

Somyadeep Shrivastava 60 Microsoft Employee

After some amount of time initially after working, I start getting 408s, which really blocks from experimentation on development. Is that a limit issue? it happens frequently User's image

Nikhil Jha (Accenture International Limited) 4,330 Reputation points Microsoft External Staff Moderator

2025-11-12T04:02:50.36+00:00

Hello Somyadeep Shrivastava,
Following up to check if you had a chance to work on suggested response in our offline discussion. If it helped kindly accept and upvote the answer.
Somyadeep Shrivastava 60 Reputation points Microsoft Employee

2025-11-13T22:48:21.26+00:00

hey, im still implementing a non-wrapper solution to bypass 408, but at the same time im getting 503 time to time

Nikhil Jha (Accenture International Limited) 4,330 Microsoft External Staff Moderator

Hello Somyadeep Shrivastava,
The 503 Service Unavailable Error: It means This is a server-side error. The Azure AI service is actively telling you that it is temporarily overloaded, at capacity, or undergoing a transient event. It is rejecting your request to protect the service.

I would recommend you try the retry with exponential backoff logic (wait 1s, then 2s, then 4s, etc.).

Sample code:

library(httr2)
# Your endpoint and key from the AI Studio deployment
API_ENDPOINT <- "YOUR_TIMEGEN1_ENDPOINT_URL"
API_KEY <- "YOUR_API_KEY"
# Your JSON body for the forecast
forecast_body <- list(
  # ... your data and parameters ...
  h = 8,
  finetune_steps = 10,
  level = c(80, 95)
)
# 1. Build the base request
req <- request(API_ENDPOINT) |>
  req_headers(
    `Authorization` = paste("Bearer", API_KEY),
    `Content-Type` = "application/json"
  ) |>
  req_body_json(forecast_body) |>
  
  # 2. FIX FOR 408: Set a 5-minute (300 sec) timeout
  req_timeout_sec(300) |> 
  # 3. FIX FOR 503: Add retry logic for transient errors
  req_retry(
    is_transient = ~ resp_status(.x) %in% c(429, 500, 503), # Retry on 503
    max_tries = 4,
    backoff = ~ 2^(.y - 1) # Exponential backoff (1s, 2s, 4s)
  )
# 4. Perform the request
tryCatch({
  resp <- req_perform(req)
  resp_data <- resp_body_json(resp)
  print("Forecast successful:")
  print(resp_data)
  
}, error = function(e) {
  print(paste("Request failed after retries:", e$message))
})

Note: This is a code samples based on available documentation along with a few custom adjustments. Since environments and requirements may vary, I would kindly recommend reviewing and validating the code in a safe or test environment before applying it to production.

Refrence:
https://learn.microsoft.com/en-us/azure/architecture/best-practices/transient-faults

Nikhil Jha (Accenture International Limited) 4,330 Reputation points Microsoft External Staff Moderator

2025-11-17T06:09:29.7833333+00:00

Hello Somyadeep Shrivastava,
I hope i provided a way ahead to your workaorund.
Could you please accept the answer and upvote helping other community members looking for similar remediation.

Answer accepted by question author

0 additional answers

Your answer

Nikhil Jha (Accenture International Limited) 4,330 Reputation points Microsoft External Staff Moderator

2025-11-12T04:02:50.36+00:00

Hello Somyadeep Shrivastava,
Following up to check if you had a chance to work on suggested response in our offline discussion. If it helped kindly accept and upvote the answer.
Somyadeep Shrivastava 60 Reputation points Microsoft Employee

2025-11-13T22:48:21.26+00:00

hey, im still implementing a non-wrapper solution to bypass 408, but at the same time im getting 503 time to time
Nikhil Jha (Accenture International Limited) 4,330 Reputation points Microsoft External Staff Moderator

2025-11-14T05:45:19.5166667+00:00

Hello Somyadeep Shrivastava,
The 503 Service Unavailable Error: It means This is a server-side error. The Azure AI service is actively telling you that it is temporarily overloaded, at capacity, or undergoing a transient event. It is rejecting your request to protect the service.

I would recommend you try the retry with exponential backoff logic (wait 1s, then 2s, then 4s, etc.).

Sample code:

library(httr2) # Your endpoint and key from the AI Studio deployment API_ENDPOINT <- "YOUR_TIMEGEN1_ENDPOINT_URL" API_KEY <- "YOUR_API_KEY" # Your JSON body for the forecast forecast_body <- list( # ... your data and parameters ... h = 8, finetune_steps = 10, level = c(80, 95) ) # 1. Build the base request req <- request(API_ENDPOINT) |> req_headers( `Authorization` = paste("Bearer", API_KEY), `Content-Type` = "application/json" ) |> req_body_json(forecast_body) |> # 2. FIX FOR 408: Set a 5-minute (300 sec) timeout req_timeout_sec(300) |> # 3. FIX FOR 503: Add retry logic for transient errors req_retry( is_transient = ~ resp_status(.x) %in% c(429, 500, 503), # Retry on 503 max_tries = 4, backoff = ~ 2^(.y - 1) # Exponential backoff (1s, 2s, 4s) ) # 4. Perform the request tryCatch({ resp <- req_perform(req) resp_data <- resp_body_json(resp) print("Forecast successful:") print(resp_data) }, error = function(e) { print(paste("Request failed after retries:", e$message)) })

Note: This is a code samples based on available documentation along with a few custom adjustments. Since environments and requirements may vary, I would kindly recommend reviewing and validating the code in a safe or test environment before applying it to production.

Refrence:
https://learn.microsoft.com/en-us/azure/architecture/best-practices/transient-faults
Nikhil Jha (Accenture International Limited) 4,330 Reputation points Microsoft External Staff Moderator

2025-11-17T06:09:29.7833333+00:00

Hello Somyadeep Shrivastava,
I hope i provided a way ahead to your workaorund.
Could you please accept the answer and upvote helping other community members looking for similar remediation.

Answer 1

Hello Somyadeep Shrivastava,

I understand you're experiencing 408 Request Timeout errors when calling the TimeGen-1 model, which is blocking your development.

This is a common issue when experimenting with new endpoints, and your suspicion is correct—it is almost certainly related to service limits.

The key detail you provided is that the service works initially and then begins to fail frequently. This pattern is a symptom of service-side throttling or rate limiting.

Here are a few recommended steps to resolve this and continue your development.

1. Implement Exponential Backoff (Retry Logic)

This is the most important step. Instead of just resending a failed request immediately, you must implement a "retry with backoff" strategy.
If a request fails:

Wait 1 second, then retry.
If it fails again, wait 2 seconds, then retry.
If it fails again, wait 4 seconds, then retry.

This pattern gives the service time to recover and gradually lets your requests through as your quota window resets.

2. Reduce Client-Side Request Frequency

Since this is happening during "experimentation," you are likely calling the model in a rapid loop or in quick succession. Manually slow down your requests. If you are running a script, add a simple time.sleep(1) or time.sleep(2) between calls to ensure you stay under the requests-per-minute limit.

3. Check for Concurrent Requests

The service also has a limit on concurrent requests (how many requests you can have "in flight" at the same time). If your code is using async or multi-threading to send many requests in parallel, you are almost certainly hitting this limit. Try sending your requests sequentially (one at a time) to see if the errors stop.

For Reference:

Please let us know if implementing the retry logic and slowing down the request frequency helps. If yes, kindly "Accept the answer" and/or upvote, so it will be beneficial to others in the community as well.

Share via

Getting 408 Request Timeout when calling TimeGen-1 timegpt

0 additional answers

Your answer